WO2019197715A1

WO2019197715A1 - An apparatus, a method and a computer program for running a neural network

Info

Publication number: WO2019197715A1
Application number: PCT/FI2019/050269
Authority: WO
Inventors: Francesco Cricri; Miska Hannuksela; Jani Lainema; Caglar AYTEKIN
Original assignee: Nokia Technologies Oy
Priority date: 2018-04-09
Filing date: 2019-04-03
Publication date: 2019-10-17
Also published as: EP3777195A1; EP3777195A4

Abstract

The present invention relates to an apparatus, a method and a computer program for running a neural network. An example method according to the invention comprises: receiving, in a first apparatus, a first portion of a video data stream at a first quality (500); using the high quality first portion of the video data stream to train a neural network for enhancing a video data stream of a type sufficiently similar to the first portion of the video data stream (502); receiving, in the first apparatus, a second portion of said video data stream at a second quality (504); and enhancing the quality of the lower quality second portion of said video data stream using the trained neural network (506).

Description

AN APPARATUS, A METHOD AND A COMPUTER PROGRAM FOR RUNNING A

NEURAL NETWORK

TECHNICAL FIELD

[0001] The present invention relates to an apparatus, a method and a computer program for running a neural network.

BACKGROUND

[0002] Recently, the development of various artificial neural network (NN) techniques, especially the ones related to deep learning, has enabled to leam algorithms for several tasks from the raw data, which algorithms may outperform algorithms which have been developed for many years using non-learning based methods.

[0003] Neural networks are used more and more in various types of devices, from smartphones to self-driving cars. Many small devices are typically very constrained in terms of memory, bandwidth and computation capacity. On the other hand, many small devices are configured to run applications, which could benefit from running the application or part of it as NN-based algorithms.

[0004] For example, neural networks based video processing may be used for enhancing a received video in the receiving device such that the device transmitting the video may send the video at lower quality. The transmitting device may comprise a video encoder and the receiving device may comprise a video decoder.

[0005] Due to high memory requirement and computationally costly implementation of typical neural networks, the training of neural networks on low capacity transmitting devices may be difficult or even impossible on the computation- limited devices. Moreover, sending a neural network and the parameters related thereto requires a wide bandwidth between the transmitting device and the receiving device.

SUMMARY

[0006] Now in order to at least alleviate the above problems, an improved method for training and running a neural network for video enhancements is introduced herein.

[0007] A method according to a first aspect comprises receiving, in a first apparatus, a first portion of a video data stream at a first quality; using the first portion of the video data stream to train a neural network for enhancing a video data stream of a type sufficiently similar to the first portion of the video data stream; receiving, in the first apparatus, a second portion of said video data stream at a second quality, wherein the second quality is lower than the first quality; and enhancing the quality of the second portion of said video data stream using the trained neural network.

[0008] According to an embodiment, the method further comprises processing the received high quality first portion of the video data stream into a lower quality; inputting the processed lower quality first portion of the video data stream in said neural network; computing a loss of the neural network on the basis of a difference between a video data stream at an output of the neural network and the received high quality first portion of the video data stream; deriving a training signal for the neural network based on the loss; and training the neural network with the training signal to decrease the loss.

[0009] According to an embodiment, the method further comprises training the neural network iteratively with an iteratively updated training signal until the loss meets at least one predetermined stopping criterion.

[0010] A method according to a second aspect comprises transmitting, by a second apparatus, a first portion of a video data stream at a first quality to a first apparatus; receiving, from the first apparatus, one or more key frames, or parts thereof, or feature vectors thereof, wherein the one or more key frames correspond to a subset of the first portion of a video data stream; and transmitting a second portion of said video data stream comprising frames sufficiently similar to the key frames at a second quality to the first apparatus, wherein the second quality is lower than the first quality.

[0011] Apparatuses according to a third and a fourth aspect comprise at least one processor and at least one memory, said at least one memory stored with code thereon, the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus to perform the method.

[0012] According to a further aspect, there is provided a computer readable non-transitory storage medium stored with code thereon for use by an apparatus, which when executed by a processor, causes the apparatus to perform the method.

[0013] The apparatuses and the computer readable storage mediums stored with code thereon, as described above, are thus arranged to carry out the above methods and one or more of the embodiments related thereto.

[0014] Further aspects relate to apparatuses comprising means for performing the above methods and one or more of the embodiments related thereto. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] For better understanding of the present invention, reference will now be made by way of example to the accompanying drawings in which:

[0016] Figure 1 shows schematically an electronic device employing embodiments of the invention;

[0017] Figure 2 shows schematically a user equipment suitable for employing

embodiments of the invention;

[0018] Figure 3 further shows schematically electronic devices employing embodiments of the invention connected using wireless and wired network connections;

[0019] Figure 4 shows schematically a block chart of an encoder on a general level;

[0020] Figure 5 shows a flow chart of a method for running a neural network according to an embodiment of the invention;

[0021] Figure 6 shows a simplified example of a setup for implementing the embodiments of the invention;

[0022] Figure 7 shows a schematic diagram of a decoder suitable for implementing embodiments of the invention;

[0023] Figure 8 illustrates the operation of a decoder for blocks/pictures selected to be enhanced by the neural net;

[0024] Figure 9 shows an illustration of how similar shots are streamed at different qualities according to an embodiment of the invention; and

[0025] Figure 10 shows a block diagram of signaling quality level selection to the transmitting device according to an embodiment of the invention.

DETAILED DESCRIPTON OF SOME EXAMPLE EMBODIMENTS

[0026] In the following, several embodiments will be described in the context of encoding and decoding visual data, such as video frames. It is to be noted, however, the embodiments are not limited to processing of visual data, but the different embodiments have applications in any environment where data can be streamed and compressed. Thus, applications including but not limited to, for example, streaming of speech or other audio data can benefit from the use of the embodiments.

[0027] The following describes in further detail suitable apparatus and possible

mechanisms for running a neural network according to embodiments. In this regard reference is first made to Figures 1 and 2, where Figure 1 shows an example block diagram of an apparatus 50. The apparatus may be an Internet of Things (IoT) apparatus configured to perform various functions, such as for example, gathering information by one or more sensors, receiving or transmitting information, analyzing information gathered or received by the apparatus, or the like. The apparatus may comprise a video coding system, which may incorporate a codec. Figure 2 shows a layout of an apparatus according to an example embodiment. The elements of Figs. 1 and 2 will be explained next.

[0028] The electronic device 50 may for example be a mobile terminal or user equipment of a wireless communication system, a sensor device, a tag, or other lower power device. However, it would be appreciated that embodiments of the invention may be implemented within any electronic device or apparatus which may process data by neural networks.

[0029] The apparatus 50 may comprise a housing 30 for incorporating and protecting the device. The apparatus 50 further may comprise a display 32 in the form of a liquid crystal display. In other embodiments of the invention the display may be any suitable display technology suitable to display an image or video. The apparatus 50 may further comprise a keypad 34. In other embodiments of the invention any suitable data or user interface mechanism may be employed. For example the user interface may be implemented as a virtual keyboard or data entry system as part of a touch-sensitive display.

[0030] The apparatus may comprise a microphone 36 or any suitable audio input which may be a digital or analogue signal input. The apparatus 50 may further comprise an audio output device which in embodiments of the invention may be any one of: an earpiece 38, speaker, or an analogue audio or digital audio output connection. The apparatus 50 may also comprise a battery (or in other embodiments of the invention the device may be powered by any suitable mobile energy device such as solar cell, fuel cell or clockwork generator). The apparatus may further comprise a camera capable of recording or capturing images and/or video. The apparatus 50 may further comprise an infrared port for short range line of sight communication to other devices. In other embodiments the apparatus 50 may further comprise any suitable short range communication solution such as for example a Bluetooth wireless connection or a USB/firewire wired connection.

[0031] The apparatus 50 may comprise a controller 56, processor or processor circuitry for controlling the apparatus 50. The controller 56 may be connected to memory 58 which in embodiments of the invention may store both data in the form of image and audio data and/or may also store instructions for implementation on the controller 56. The controller 56 may further be connected to codec circuitry 54 suitable for carrying out coding and/or decoding of audio and/or video data or assisting in coding and/or decoding carried out by the controller. [0032] The apparatus 50 may further comprise a card reader 48 and a smart card 46, for example a UICC and UICC reader for providing user information and being suitable for providing authentication information for authentication and authorization of the user at a network.

[0033] The apparatus 50 may comprise radio interface circuitry 52 connected to the controller and suitable for generating wireless communication signals for example for communication with a cellular communications network, a wireless communications system or a wireless local area network. The apparatus 50 may further comprise an antenna 44 connected to the radio interface circuitry 52 for transmitting radio frequency signals generated at the radio interface circuitry 52 to other apparatus(es) and/or for receiving radio frequency signals from other apparatus(es).

[0034] The apparatus 50 may comprise a camera capable of recording or detecting individual frames which are then passed to the codec 54 or the controller for processing. The apparatus may receive the video image data for processing from another device prior to transmission and/or storage. The apparatus 50 may also receive either wirelessly or by a wired connection the image for coding/decoding. The structural elements of apparatus 50 described above represent examples of means for performing a corresponding function.

[0035] With respect to Figure 3, an example of a system within which embodiments of the present invention can be utilized is shown. The system 10 comprises multiple communication devices which can communicate through one or more networks. The system 10 may comprise any combination of wired or wireless networks including, but not limited to a wireless cellular telephone network (such as a GSM, UMTS, CDMA, 4G, 5G network etc.), a wireless local area network (WLAN) such as defined by any of the IEEE 802.x standards, a Bluetooth personal area network, an Ethernet local area network, a token ring local area network, a wide area network, and the Internet.

[0036] The system 10 may include both wired and wireless communication devices and/or apparatus 50 suitable for implementing embodiments of the invention.

[0037] For example, the system shown in Figure 3 shows a mobile telephone network 11 and a representation of the internet 28. Connectivity to the internet 28 may include, but is not limited to, long range wireless connections, short range wireless connections, and various wired connections including, but not limited to, telephone lines, cable lines, power lines, and similar communication pathways.

[0038] The example communication devices shown in the system 10 may include, but are not limited to, an electronic device or apparatus 50, a combination of a personal digital assistant (PDA) and a mobile telephone 14, a PDA 16, an integrated messaging device (IMD) 18, a desktop computer 20, a notebook computer 22. The apparatus 50 may be stationary or mobile when carried by an individual who is moving. The apparatus 50 may also be located in a mode of transport including, but not limited to, a car, a truck, a taxi, a bus, a train, a boat, an airplane, a bicycle, a motorcycle or any similar suitable mode of transport.

[0039] The embodiments may also be implemented in a set-top box; i.e. a digital TV receiver, which may/may not have a display or wireless capabilities, in tablets or (laptop) personal computers (PC), which have hardware and/or software to process neural network data, in various operating systems, and in chipsets, processors, DSPs and/or embedded systems offering hardware/software based coding.

[0040] Some or further apparatus may send and receive calls and messages and communicate with service providers through a wireless connection 25 to a base station 24.

The base station 24 may be connected to a network server 26 that allows communication between the mobile telephone network 11 and the internet 28. The system may include additional communication devices and communication devices of various types.

[0041] The communication devices may communicate using various transmission technologies including, but not limited to, code division multiple access (CDMA), global systems for mobile communications (GSM), universal mobile telecommunications system (UMTS), time divisional multiple access (TDMA), frequency division multiple access (FDMA), transmission control protocol-internet protocol (TCP-IP), short messaging service (SMS), multimedia messaging service (MMS), email, instant messaging service (IMS), Bluetooth, IEEE 802.11, 3GPP Narrowband IoT and any similar wireless communication technology. A communications device involved in implementing various embodiments of the present invention may communicate using various media including, but not limited to, radio, infrared, laser, cable connections, and any suitable connection.

[0042] In telecommunications and data networks, a channel may refer either to a physical channel or to a logical channel. A physical channel may refer to a physical transmission medium such as a wire, whereas a logical channel may refer to a logical connection over a multiplexed medium, capable of conveying several logical channels. A channel may be used for conveying an information signal, for example a bitstream, from one or several senders (or transmitters) to one or several receivers.

[0043] The embodiments may also be implemented in so-called IoT devices. The Internet of Things (IoT) may be defined, for example, as an interconnection of uniquely identifiable embedded computing devices within the existing Internet infrastructure. The convergence of various technologies has and will enable many fields of embedded systems, such as wireless sensor networks, control systems, home/building automation, etc. to be included the Internet of Things (IoT). In order to utilize Internet IoT devices are provided with an IP address as a unique identifier. IoT devices may be provided with a radio transmitter, such as WLAN or Bluetooth transmitter or a RFID tag. Alternatively, IoT devices may have access to an IP- based network via a wired network, such as an Ethernet-based network or a power-line connection (PLC).

[0044] An MPEG-2 transport stream (TS), specified in ISO/IEC 13818-1 or equivalently in ITU-T Recommendation H.222.0, is a format for carrying audio, video, and other media as well as program metadata or other metadata, in a multiplexed stream. A packet identifier (PID) is used to identify an elementary stream (a.k.a. packetized elementary stream) within the TS. Hence, a logical channel within an MPEG-2 TS may be considered to correspond to a specific PID value.

[0045] Available media file format standards include ISO base media file format (ISO/IEC 14496-12, which may be abbreviated ISOBMFF) and file format for NAL unit structured video (ISO/IEC 14496-15), which derives from the ISOBMFF.

[0046] Video codec consists of an encoder that transforms the input video into a compressed representation suited for storage/transmission and a decoder that can uncompress the compressed video representation back into a viewable form. A video encoder and/or a video decoder may also be separate from each other, i.e. need not form a codec. Typically encoder discards some information in the original video sequence in order to represent the video in a more compact form (that is, at lower bitrate).

[0047] Typical hybrid video encoders, for example many encoder implementations of ITU- T H.263 and H.264, encode the video information in two phases. Firstly pixel values in a certain picture area (or“block”) are predicted for example by motion compensation means (finding and indicating an area in one of the previously coded video frames that corresponds closely to the block being coded) or by spatial means (using the pixel values around the block to be coded in a specified manner). Secondly the prediction error, i.e. the difference between the predicted block of pixels and the original block of pixels, is coded. This is typically done by transforming the difference in pixel values using a specified transform (e.g. Discrete Cosine Transform (DCT) or a variant of it), quantizing the coefficients and entropy coding the quantized coefficients. By varying the fidelity of the quantization process, encoder can control the balance between the accuracy of the pixel representation (picture quality) and size of the resulting coded video representation (file size or transmission bitrate). [0048] In temporal prediction, the sources of prediction are previously decoded pictures (a.k.a. reference pictures). In intra block copy (IBC; a.k.a. intra-block-copy prediction), prediction is applied similarly to temporal prediction but the reference picture is the current picture and only previously decoded samples can be referred in the prediction process. Inter layer or inter- view prediction may be applied similarly to temporal prediction, but the reference picture is a decoded picture from another scalable layer or from another view, respectively. In some cases, inter prediction may refer to temporal prediction only, while in other cases inter prediction may refer collectively to temporal prediction and any of intra block copy, inter-layer prediction, and inter- view prediction provided that they are performed with the same or similar process than temporal prediction. Inter prediction or temporal prediction may sometimes be referred to as motion compensation or motion-compensated prediction.

[0049] Inter prediction, which may also be referred to as temporal prediction, motion compensation, or motion-compensated prediction, reduces temporal redundancy. In inter prediction the sources of prediction are previously decoded pictures. Intra prediction utilizes the fact that adjacent pixels within the same picture are likely to be correlated. Intra prediction can be performed in spatial or transform domain, i.e., either sample values or transform coefficients can be predicted. Intra prediction is typically exploited in intra coding, where no inter prediction is applied.

[0050] One outcome of the coding procedure is a set of coding parameters, such as motion vectors and quantized transform coefficients. Many parameters can be entropy-coded more efficiently if they are predicted first from spatially or temporally neighboring parameters. For example, a motion vector may be predicted from spatially adjacent motion vectors and only the difference relative to the motion vector predictor may be coded. Prediction of coding parameters and intra prediction may be collectively referred to as in-picture prediction.

[0051 ] Figure 4 shows a block diagram of a general structure of a video encoder. Figure 4 presents an encoder for two layers, but it would be appreciated that presented encoder could be similarly extended to encode more than two layers. Figure 4 illustrates a video encoder comprising a first encoder section 500 for a base layer and a second encoder section 502 for an enhancement layer. Each of the first encoder section 500 and the second encoder section 502 may comprise similar elements for encoding incoming pictures. The encoder sections 500, 502 may comprise a pixel predictor 302, 402, prediction error encoder 303, 403 and prediction error decoder 304, 404. Figure 4 also shows an embodiment of the pixel predictor 302, 402 as comprising an inter-predictor 306, 406, an intra-predictor 308, 408, a mode selector 310, 410, a filter 316, 416, and a reference frame memory 318, 418. The pixel predictor 302 of the first encoder section 500 receives 300 base layer images of a video stream to be encoded at both the inter-predictor 306 (which determines the difference between the image and a motion compensated reference frame 318) and the intra-predictor 308 (which determines a prediction for an image block based only on the already processed parts of current frame or picture). The output of both the inter-predictor and the intra-predictor are passed to the mode selector 310. The intra-predictor 308 may have more than one intra prediction modes. Hence, each mode may perform the intra-prediction and provide the predicted signal to the mode selector 310. The mode selector 310 also receives a copy of the base layer picture 300. Correspondingly, the pixel predictor 402 of the second encoder section 502 receives 400 enhancement layer images of a video stream to be encoded at both the inter predictor 406 (which determines the difference between the image and a motion compensated reference frame 418) and the intra-predictor 408 (which determines a prediction for an image block based only on the already processed parts of current frame or picture). The output of both the inter-predictor and the intra-predictor are passed to the mode selector 410. The intra predictor 408 may have more than one intra-prediction modes. Hence, each mode may perform the intra-prediction and provide the predicted signal to the mode selector 410. The mode selector 410 also receives a copy of the enhancement layer picture 400.

[0052] Depending on which encoding mode is selected to encode the current block, the output of the inter-predictor 306, 406 or the output of one of the optional intra-predictor modes or the output of a surface encoder within the mode selector is passed to the output of the mode selector 310, 410. The output of the mode selector is passed to a first summing device 321, 421. The first summing device may subtract the output of the pixel predictor 302, 402 from the base layer picture 300/enhancement layer picture 400 to produce a first prediction error signal 320, 420 which is input to the prediction error encoder 303, 403.

[0053] The pixel predictor 302, 402 further receives from a preliminary reconstructor 339, 439 the combination of the prediction representation of the image block 312, 412 and the output 338, 438 of the prediction error decoder 304, 404. The preliminary reconstructed image 314, 414 may be passed to the intra-predictor 308, 408 and to a filter 316, 416. The filter 316, 416 receiving the preliminary representation may filter the preliminary

representation and output a final reconstructed image 340, 440 which may be saved in a reference frame memory 318, 418. The reference frame memory 318 may be connected to the inter-predictor 306 to be used as the reference image against which a future base layer picture 300 is compared in inter-prediction operations. Subject to the base layer being selected and indicated to be source for inter-layer sample prediction and/or inter-layer motion information prediction of the enhancement layer according to some embodiments, the reference frame memory 318 may also be connected to the inter-predictor 406 to be used as the reference image against which a future enhancement layer pictures 400 is compared in inter-prediction operations. Moreover, the reference frame memory 418 may be connected to the inter predictor 406 to be used as the reference image against which a future enhancement layer picture 400 is compared in inter-prediction operations.

[0054] Filtering parameters from the filter 316 of the first encoder section 500 may be provided to the second encoder section 502 subject to the base layer being selected and indicated to be source for predicting the filtering parameters of the enhancement layer according to some embodiments.

[0055] The prediction error encoder 303, 403 comprises a transform unit 342, 442 and a quantizer 344, 444. The transform unit 342, 442 transforms the first prediction error signal 320, 420 to a transform domain. The transform is, for example, the DCT transform. The quantizer 344, 444 quantizes the transform domain signal, e.g. the DCT coefficients, to form quantized coefficients.

[0056] The prediction error decoder 304, 404 receives the output from the prediction error encoder 303, 403 and performs the opposite processes of the prediction error encoder 303,

403 to produce a decoded prediction error signal 338, 438 which, when combined with the prediction representation of the image block 312, 412 at the second summing device 339, 439, produces the preliminary reconstructed image 314, 414. The prediction error decoder may be considered to comprise a dequantizer 361, 461, which dequantizes the quantized coefficient values, e.g. DCT coefficients, to reconstruct the transform signal and an inverse

transformation unit 363, 463, which performs the inverse transformation to the reconstructed transform signal wherein the output of the inverse transformation unit 363, 463 contains reconstructed block(s). The prediction error decoder may also comprise a block filter which may filter the reconstructed block(s) according to further decoded information and filter parameters.

[0057] The entropy encoder 330, 430 receives the output of the prediction error encoder 303, 403 and may perform a suitable entropy encoding/variable length encoding on the signal to provide error detection and correction capability. The outputs of the entropy encoders 330, 430 may be inserted into a bitstream e.g. by a multiplexer 508.

[0058] Entropy coding/decoding may be performed in many ways. For example, context- based coding/decoding may be applied, where in both the encoder and the decoder modify the context state of a coding parameter based on previously coded/decoded coding parameters. Context-based coding may for example be context adaptive binary arithmetic coding

(CABAC) or context-based variable length coding (CAVLC) or any similar entropy coding. Entropy coding/decoding may alternatively or additionally be performed using a variable length coding scheme, such as Huffman coding/decoding or Exp-Golomb coding/decoding. Decoding of coding parameters from an entropy-coded bitstream or codewords may be referred to as parsing.

[0059] The H.264/AVC standard was developed by the Joint Video Team (JVT) of the Video Coding Experts Group (VCEG) of the Telecommunications Standardization Sector of International Telecommunication Union (ITU-T) and the Moving Picture Experts Group (MPEG) of International Organisation for Standardization (ISO) / International

Electrotechnical Commission (IEC). The H.264/AVC standard is published by both parent standardization organizations, and it is referred to as ITU-T Recommendation H.264 and ISO/IEC International Standard 14496-10, also known as MPEG-4 Part 10 Advanced Video Coding (AVC). There have been multiple versions of the H.264/ AVC standard, integrating new extensions or features to the specification. These extensions include Scalable Video Coding (SVC) and Multiview Video Coding (MVC).

[0060] Version 1 of the High Efficiency Video Coding (H.265/HEVC a.k.a. HEVC) standard was developed by the Joint Collaborative Team - Video Coding (JCT-VC) of VCEG and MPEG. The standard was published by both parent standardization organizations, and it is referred to as ITU-T Recommendation H.265 and ISO/IEC International Standard 23008-2, also known as MPEG-H Part 2 High Efficiency Video Coding (HEVC). Later versions of H.265/HEVC included scalable, multiview, fidelity range extensions, , three-dimensional, and screen content coding extensions which may be abbreviated SHVC, MV-HEVC, REXT, 3D- HEVC, and SCC, respectively.

[0061] SHVC, MV-HEVC, and 3D-HEVC use a common basis specification, specified in Annex F of the version 2 of the HEVC standard. This common basis comprises for example high-level syntax and semantics e.g. specifying some of the characteristics of the layers of the bitstream, such as inter-layer dependencies, as well as decoding processes, such as reference picture list construction including inter-layer reference pictures and picture order count derivation for multi-layer bitstream. Annex F may also be used in potential subsequent multi layer extensions of HEVC. It is to be understood that even though a video encoder, a video decoder, encoding methods, decoding methods, bitstream structures, and/or embodiments may be described in the following with reference to specific extensions, such as SHVC and/or MV-HEVC, they are generally applicable to any multi-layer extensions of HEVC, and even more generally to any multi-layer video coding scheme.

[0062] Some key definitions, bitstream and coding structures, and concepts of H.264/AVC and HEVC are described in this section as an example of a video encoder, decoder, encoding method, decoding method, and a bitstream structure, wherein the embodiments may be implemented. Some of the key definitions, bitstream and coding structures, and concepts of H.264/AVC are the same as in HEVC - hence, they are described below jointly. The aspects of the invention are not limited to H.264/AVC or HEVC, but rather the description is given for one possible basis on top of which the invention may be partly or fully realized.

[0063] Similarly to many earlier video coding standards, the bitstream syntax and semantics as well as the decoding process for error-free bitstreams are specified in

H.264/AVC and HEVC. The encoding process is not specified, but encoders must generate conforming bitstreams. Bitstream and decoder conformance can be verified with the

Hypothetical Reference Decoder (HRD). The standards contain coding tools that help in coping with transmission errors and losses, but the use of the tools in encoding is optional and no decoding process has been specified for erroneous bitstreams.

[0064] The elementary unit for the input to an H.264/AVC or HEVC encoder and the output of an H.264/AVC or HEVC decoder, respectively, is a picture. A picture given as an input to an encoder may also be referred to as a source picture, and a picture decoded by a decoded may be referred to as a decoded picture.

[0065] The source and decoded pictures are each comprised of one or more sample arrays, such as one of the following sets of sample arrays:

[0066] Luma (Y) only (monochrome).

[0067] Luma and two chroma (YCbCr or YCgCo).

[0068] Green, Blue and Red (GBR, also known as RGB).

[0069] Arrays representing other unspecified monochrome or tri- stimulus color samplings (for example, YZX, also known as XYZ).

[0070] In the following, these arrays may be referred to as luma (or L or Y) and chroma, where the two chroma arrays may be referred to as Cb and Cr; regardless of the actual color representation method in use. The actual color representation method in use can be indicated e.g. in a coded bitstream e.g. using the Video Usability Information (VUI) syntax of

H.264/AVC and/or HEVC. A component may be defined as an array or single sample from one of the three sample arrays (luma and two chroma) or the array or a single sample of the array that compose a picture in monochrome format. [0071] In H.264/AVC and HEVC, a picture may either be a frame or a field. A frame comprises a matrix of luma samples and possibly the corresponding chroma samples. A field is a set of alternate sample rows of a frame and may be used as encoder input, when the source signal is interlaced. Chroma sample arrays may be absent (and hence monochrome sampling may be in use) or chroma sample arrays may be subsampled when compared to luma sample arrays. Chroma formats may be summarized as follows:

[0072] In monochrome sampling there is only one sample array, which may be nominally considered the luma array.

[0073] In 4:2:0 sampling, each of the two chroma arrays has half the height and half the width of the luma array.

[0074] In 4:2:2 sampling, each of the two chroma arrays has the same height and half the width of the luma array.

[0075] In 4:4:4 sampling when no separate color planes are in use, each of the two chroma arrays has the same height and width as the luma array.

[0076] In H.264/AVC and HEVC, it is possible to code sample arrays as separate color planes into the bitstream and respectively decode separately coded color planes from the bitstream. When separate color planes are in use, each one of them is separately processed (by the encoder and/or the decoder) as a picture with monochrome sampling.

[0077] A partitioning may be defined as a division of a set into subsets such that each element of the set is in exactly one of the subsets.

[0078] When describing the operation of HEVC encoding and/or decoding, the following terms may be used. A coding block may be defined as an NxN block of samples for some value of N such that the division of a coding tree block into coding blocks is a partitioning. A coding tree block (CTB) may be defined as an NxN block of samples for some value of N such that the division of a component into coding tree blocks is a partitioning. A coding tree unit (CTU) may be defined as a coding tree block of luma samples, two corresponding coding tree blocks of chroma samples of a picture that has three sample arrays, or a coding tree block of samples of a monochrome picture or a picture that is coded using three separate color planes and syntax structures used to code the samples. A coding unit (CU) may be defined as a coding block of luma samples, two corresponding coding blocks of chroma samples of a picture that has three sample arrays, or a coding block of samples of a monochrome picture or a picture that is coded using three separate color planes and syntax structures used to code the samples. A CU with the maximum allowed size may be named as LCU (largest coding unit) or coding tree unit (CTU) and the video picture is divided into non-overlapping LCUs. [0079] A CU consists of one or more prediction units (PU) defining the prediction process for the samples within the CU and one or more transform units (TU) defining the prediction error coding process for the samples in the said CU. Typically, a CU consists of a square block of samples with a size selectable from a predefined set of possible CU sizes. Each PU and TU can be further split into smaller PUs and TUs in order to increase granularity of the prediction and prediction error coding processes, respectively. Each PU has prediction information associated with it defining what kind of a prediction is to be applied for the pixels within that PU (e.g. motion vector information for inter predicted PUs and intra prediction directionality information for intra predicted PUs).

[0080] Each TU can be associated with information describing the prediction error decoding process for the samples within the said TU (including e.g. DCT coefficient information). It is typically signalled at CU level whether prediction error coding is applied or not for each CU. In the case there is no prediction error residual associated with the CU, it can be considered there are no TUs for the said CU. The division of the image into CUs, and division of CUs into PUs and TUs is typically signalled in the bitstream allowing the decoder to reproduce the intended structure of these units.

[0081 ] In some coding systems, such as HEVC, a picture can be partitioned in tiles, which are rectangular and contain an integer number of blocks, such as LCUs in HEVC. In HEVC, the partitioning to tiles forms a regular grid that may be characterized by a list of tile column widths and a list of tile row heights. Tiles are ordered in the bitstream consecutively in the raster scan order of the tile grid. A tile may contain an integer number of slices, or a slice may contain an integer number of tiles. The blocks (such as CTUs in HEVC) may be scanned in encoding and decoding tile-wise in the raster scan order of blocks, and tiles may be scanned in raster scan order along the tile grid.

[0082] Video coding standards and specifications may allow encoders to divide a coded picture to coded slices or alike. In-picture prediction is typically disabled across slice boundaries. Thus, slices can be regarded as a way to split a coded picture to independently decodable pieces. Slices are therefore often regarded as elementary units for transmission. In many cases, encoders may indicate in the bitstream which types of in-picture prediction are turned off across slice boundaries, and the decoder operation takes this information into account for example when concluding which prediction sources are available. For example, samples from a neighbouring block may be regarded as unavailable for intra prediction, if the neighbouring block resides in a different slice. [0083] In HEVC, a slice is defined to be an integer number of coding tree units contained in one independent slice segment and all subsequent dependent slice segments (if any) that precede the next independent slice segment (if any) within the same access unit. In HEVC, a slice segment is defined to be an integer number of coding tree units ordered consecutively in the tile scan and contained in a single NAL unit. The division of each picture into slice segments is a partitioning. In HEVC, an independent slice segment is defined to be a slice segment for which the values of the syntax elements of the slice segment header are not inferred from the values for a preceding slice segment, and a dependent slice segment is defined to be a slice segment for which the values of some syntax elements of the slice segment header are inferred from the values for the preceding independent slice segment in decoding order. In HEVC, a slice header is defined to be the slice segment header of the independent slice segment that is a current slice segment or is the independent slice segment that precedes a current dependent slice segment, and a slice segment header is defined to be a part of a coded slice segment containing the data elements pertaining to the first or all coding tree units represented in the slice segment. The CUs are scanned in the raster scan order of LCUs within tiles or within a picture, if tiles are not in use. Within an LCU, the CUs have a specific scan order.

[0084] In wavefront parallel processing (WPP) each block row (such as CTU row in HEVC) of a slice can be encoded and decoded in parallel. When WPP is used, the state of the entropy codec at the beginning of a block row is obtained from the state of the entropy codec of the block row above after processing the second block of that row. Consequently, block rows can be processed in parallel with a delay of 2 blocks per each block row.

[0085] A motion-constrained tile set (MCTS) is such that the inter prediction process is constrained in encoding such that no sample value outside the motion-constrained tile set, and no sample value at a fractional sample position that is derived using one or more sample values outside the motion-constrained tile set, is used for inter prediction of any sample within the motion-constrained tile set. Additionally, the encoding of an MCTS is constrained in a manner that motion vector candidates are not derived from blocks outside the MCTS.

This may be enforced by turning off temporal motion vector prediction of HEVC, or by disallowing the encoder to use the TMVP candidate or any motion vector prediction candidate following the TMVP candidate in the merge or AMVP candidate list for PUs located directly left of the right tile boundary of the MCTS except the last one at the bottom right of the MCTS. In general, an MCTS may be defined to be a tile set that is independent of any sample values and coded data, such as motion vectors, that are outside the MCTS. In some cases, an MCTS may be required to form a rectangular area. It should be understood that depending on the context, an MCTS may refer to the tile set within a picture or to the respective tile set in a sequence of pictures. The respective tile set may be, but in general need not be, collocated in the sequence of pictures.

[0086] It is noted that slices, wavefronts and tiles, also when included in MCTSs, may be used as parallelization tools to enable running several encoding and/or decoding threads or processes in parallel.

[0087] It is noted that sample locations used in inter prediction may be saturated by the encoding and/or decoding process so that a location that would be outside the picture otherwise is saturated to point to the corresponding boundary sample of the picture. Hence, if a tile boundary is also a picture boundary, in some use cases, encoders may allow motion vectors to effectively cross that boundary or a motion vector to effectively cause fractional sample interpolation that would refer to a location outside that boundary, since the sample locations are saturated onto the boundary. In other use cases, specifically if a coded tile may be extracted from a bitstream where it is located on a position adjacent to a picture boundary to another bitstream where the tile is located on a position that is not adjacent to a picture boundary, encoders may constrain the motion vectors on picture boundaries similarly to any MCTS boundaries.

[0088] The temporal motion-constrained tile sets SEI message of HEVC can be used to indicate the presence of motion-constrained tile sets in the bitstream.

[0089] The decoder reconstructs the output video by applying prediction means similar to the encoder to form a predicted representation of the pixel blocks (using the motion or spatial information created by the encoder and stored in the compressed representation) and prediction error decoding (inverse operation of the prediction error coding recovering the quantized prediction error signal in spatial pixel domain). After applying prediction and prediction error decoding means the decoder sums up the prediction and prediction error signals (pixel values) to form the output video frame. The decoder (and encoder) can also apply additional filtering means to improve the quality of the output video before passing it for display and/or storing it as prediction reference for the forthcoming frames in the video sequence.

[0090] The filtering may for example include one more of the following: deblocking, sample adaptive offset (SAO), and/or adaptive loop filtering (ALF). H.264/AVC includes a deblocking, whereas HEVC includes both deblocking and SAO. [0091] The deblocking loop filter may include multiple filtering modes or strengths, which may be adaptively selected based on the features of the blocks adjacent to the boundary, such as the quantization parameter value, and/or signaling included by the encoder in the bitstream. For example, the deblocking loop filter may comprise a normal filtering mode and a strong filtering mode, which may differ in terms of the number of filter taps (i.e. number of samples being filtered on both sides of the boundary) and/or the filter tap values. For example, filtering of two samples along both sides of the boundary may be performed with a filter having the impulse response of (3 7 9 -3)/l 6, when omitting the potential impact of a clipping operation.

[0092] An example of SAO is given next with reference to HEVC; however, SAO can be similarly applied to other coding schemes too. In SAO, a picture is divided into regions where a separate SAO decision is made for each region. In HEVC, the basic unit for adapting SAO parameters is CTU (therefore an SAO region is the block covered by the corresponding CTU).

[0093] In the SAO algorithm, samples in a CTU are classified according to a set of rules and each classified set of samples are enhanced by adding offset values. The offset values are signalled in the bitstream. There are two types of offsets: 1) Band offset 2) Edge offset. For a CTU, either no SAO or band offset or edge offset is employed. Choice of whether no SAO or band or edge offset to be used may be decided by the encoder with e.g. rate distortion optimization (RDO) and signaled to the decoder.

[0094] In the band offset, the whole range of sample values is in some embodiments divided into 32 equal-width bands. For example, for 8-bit samples, width of a band is 8 (=256/32). Out of 32 bands, 4 of them are selected and different offsets are signalled for each of the selected bands. The selection decision is made by the encoder and may be signalled as follows: The index of the first band is signalled and then it is inferred that the following four bands are the chosen ones. The band offset may be useful in correcting errors in smooth regions.

[0095] In the edge offset type, the edge offset (EO) type may be chosen out of four possible types (or edge classifications) where each type is associated with a direction: 1) vertical, 2) horizontal, 3) 135 degrees diagonal, and 4) 45 degrees diagonal. The choice of the direction is given by the encoder and signalled to the decoder. Each type defines the location of two neighbour samples for a given sample based on the angle. Then each sample in the CTU is classified into one of five categories based on comparison of the sample value against the values of the two neighbour samples. After each sample in an edge offset type CTU is classified as one of the five categories, an offset value for each of the first four categories is determined and signalled to the decoder. The offset for each category is added to the sample values associated with the corresponding category. Edge offsets may be effective in correcting ringing artifacts.

[0096] The adaptive loop filter (ALF) is another method to enhance quality of the reconstructed samples. This may be achieved by filtering the sample values in the loop. In some embodiments the encoder determines which region of the pictures are to be filtered and the filter coefficients based on e.g. RDO and this information is signalled to the decoder.

[0097] In typical video codecs the motion information is indicated with motion vectors associated with each motion compensated image block, such as a prediction unit. Each of these motion vectors represents the displacement of the image block in the picture to be coded (in the encoder side) or decoded (in the decoder side) and the prediction source block in one of the previously coded or decoded pictures. In order to represent motion vectors efficiently those are typically coded differentially with respect to block specific predicted motion vectors. In typical video codecs the predicted motion vectors are created in a predefined way, for example calculating the median of the encoded or decoded motion vectors of the adjacent blocks. Another way to create motion vector predictions is to generate a list of candidate predictions from adjacent blocks and/or co-located blocks in temporal reference pictures and signalling the chosen candidate as the motion vector predictor. In addition to predicting the motion vector values, it can be predicted which reference picture(s) are used for motion- compensated prediction and this prediction information may be represented for example by a reference index of previously coded/decoded picture. The reference index is typically predicted from adjacent blocks and/or co-located blocks in temporal reference picture.

Moreover, typical high efficiency video codecs employ an additional motion information coding/decoding mechanism, often called merging/merge mode, where all the motion field information, which includes motion vector and corresponding reference picture index for each available reference picture list, is predicted and used without any modification/correction. Similarly, predicting the motion field information is carried out using the motion field information of adjacent blocks and/or co-located blocks in temporal reference pictures and the used motion field information is signalled among a list of motion field candidate list filled with motion field information of available adjacent/co-located blocks.

[0098] In typical video codecs the prediction residual after motion compensation is first transformed with a transform kernel (like DCT) and then coded. The reason for this is that often there still exists some correlation among the residual and transform can in many cases help reduce this correlation and provide more efficient coding. [0099] Typical video encoders utilize Lagrangian cost functions to find optimal coding modes, e.g. the desired coding mode for a block and associated motion vectors. This kind of cost function uses a weighting factor l to tie together the (exact or estimated) image distortion due to lossy coding methods and the (exact or estimated) amount of information that is required to represent the pixel values in an image area:

[0100]

[0101] C = D + kR, (1)

[0102]

[0103] where C is the Lagrangian cost to be minimized, D is the image distortion (e.g. Mean Squared Error) with the mode and motion vectors considered, and R the number of bits needed to represent the required data to reconstruct the image block in the decoder (including the amount of data to represent the candidate motion vectors).

[0104] Video coding standards and specifications may allow encoders to divide a coded picture to coded slices or alike. In-picture prediction is typically disabled across slice boundaries. Thus, slices can be regarded as a way to split a coded picture to independently decodable pieces. In H.264/AVC and HEVC, in-picture prediction may be disabled across slice boundaries. Thus, slices can be regarded as a way to split a coded picture into independently decodable pieces, and slices are therefore often regarded as elementary units for transmission. In many cases, encoders may indicate in the bitstream which types of in picture prediction are turned off across slice boundaries, and the decoder operation takes this information into account for example when concluding which prediction sources are available. For example, samples from a neighboring CU may be regarded as unavailable for intra prediction, if the neighboring CU resides in a different slice.

[0105] An elementary unit for the output of an H.264/AVC or HEVC encoder and the input of an H.264/AVC or HEVC decoder, respectively, is a Network Abstraction Layer (NAL) unit. For transport over packet-oriented networks or storage into structured files, NAL units may be encapsulated into packets or similar structures. A bytestream format has been specified in H.264/AVC and HEVC for transmission or storage environments that do not provide framing structures. The bytestream format separates NAL units from each other by attaching a start code in front of each NAL unit. To avoid false detection of NAL unit boundaries, encoders run a byte-oriented start code emulation prevention algorithm, which adds an emulation prevention byte to the NAL unit payload if a start code would have occurred otherwise. In order to enable straightforward gateway operation between packet- and stream-oriented systems, start code emulation prevention may always be performed regardless of whether the bytestream format is in use or not. A NAL unit may be defined as a syntax structure containing an indication of the type of data to follow and bytes containing that data in the form of an RBSP interspersed as necessary with emulation prevention bytes. A raw byte sequence payload (RBSP) may be defined as a syntax structure containing an integer number of bytes that is encapsulated in a NAL unit. An RBSP is either empty or has the form of a string of data bits containing syntax elements followed by an RBSP stop bit and followed by zero or more subsequent bits equal to 0.

[0106] NAL units consist of a header and payload. In H.264/AVC and HEVC, the NAL unit header indicates the type of the NAL unit

[0107] In HEVC, a two-byte NAL unit header is used for all specified NAL unit types. The NAL unit header contains one reserved bit, a six-bit NAL unit type indication, a three-bit nuh_temporal_id_plusl indication for temporal level (may be required to be greater than or equal to 1) and a six-bit nuh layer id syntax element. The temporal_id_plusl syntax element may be regarded as a temporal identifier for the NAL unit, and a zero-based Temporalld variable may be derived as follows: Temporalld = temporal_id_plusl - 1. The abbreviation TID may be used to interchangeably with the Temporalld variable. Temporalld equal to 0 corresponds to the lowest temporal level. The value of temporal_id_plusl is required to be non-zero in order to avoid start code emulation involving the two NAL unit header bytes. The bitstream created by excluding all VCL NAL units having a Temporalld greater than or equal to a selected value and including all other VCL NAL units remains conforming.

Consequently, a picture having Temporalld equal to tid value does not use any picture having a Temporalld greater than tid value as inter prediction reference. A sub-layer or a temporal sub-layer may be defined to be a temporal scalable layer (or a temporal layer, TL) of a temporal scalable bitstream, consisting of VCL NAL units with a particular value of the Temporalld variable and the associated non-VCL NAL units. A picture at temporal sub-layer with Temporalld equal to N may be predicted from reference pictures at temporal sub-layers with Temporalld less than or equal to N (unless further constrained by the picture type as explained below) and is disallowed or disabled to be predicted from reference pictures at temporal sub-layers with Temporalld greater than N. nuh layer id can be understood as a scalability layer identifier.

[0108] NAL units can be categorized into Video Coding Layer (VCL) NAL units and non- VCL NAL units. VCL NAL units are typically coded slice NAL units. In HEVC, VCL NAL units contain syntax elements representing one or more CU. [0109] In HEVC, abbreviations for picture types may be defined as follows: trailing (TRAIL) picture, Temporal Sub-layer Access (TSA), Step-wise Temporal Sub-layer Access (STSA), Random Access Decodable Leading (RADL) picture, Random Access Skipped Leading (RASL) picture, Broken Link Access (BLA) picture, Instantaneous Decoding Refresh (IDR) picture, Clean Random Access (CRA) picture.

[01 10] A Random Access Point (RAP) picture, which may also be referred to as an intra random access point (IRAP) picture in an independent layer contains only intra-coded slices. An IRAP picture belonging to a predicted layer may contain P, B, and I slices, cannot use inter prediction from other picturesin the same predicted layer, and may use inter-layer prediction from its direct reference layers. In the present version of HEVC, an IRAP picture may be a BLA picture, a CRA picture or an IDR picture. The first picture in a bitstream containing a base layer is an IRAP picture at the base layer. Provided the necessary parameter sets are available when they need to be activated, an IRAP picture at an independent layer and all subsequent non-RASL pictures at the independent layer in decoding order can be correctly decoded without performing the decoding process of any pictures that precede the IRAP picture in decoding order. The IRAP picture belonging to a predicted layer and all subsequent non-RASL pictures in decoding order within the same predicted layer can be correctly decoded without performing the decoding process of any pictures of the same predicted layer that precede the IRAP picture in decoding order, when the necessary parameter sets are available when they need to be activated and when the decoding of each direct reference layer of the predicted layer has been initialized . There may be pictures in a bitstream that contain only intra-coded slices that are not IRAP pictures.

[0111] In HEVC there are two picture types, the TSA and STSA picture types that can be used to indicate temporal sub-layer switching points. If temporal sub-layers with Temporalld up to N had been decoded until the TSA or STSA picture (exclusive) and the TSA or STSA picture has Temporalld equal to N+l, the TSA or STSA picture enables decoding of all subsequent pictures (in decoding order) having Temporalld equal to N+l. The TSA picture type may impose restrictions on the TSA picture itself and all pictures in the same sub-layer that follow the TSA picture in decoding order. None of these pictures is allowed to use inter prediction from any picture in the same sub-layer that precedes the TSA picture in decoding order. The TSA definition may further impose restrictions on the pictures in higher sub-layers that follow the TSA picture in decoding order. None of these pictures is allowed to refer a picture that precedes the TSA picture in decoding order if that picture belongs to the same or higher sub-layer as the TSA picture. TSA pictures have Temporalld greater than 0. The STSA is similar to the TSA picture but does not impose restrictions on the pictures in higher sub- layers that follow the STSA picture in decoding order and hence enable up-switching only onto the sub-layer where the STSA picture resides.

[0112] A non-VCL NAL unit may be for example one of the following types: a sequence parameter set, a picture parameter set, a supplemental enhancement information (SEI) NAL unit, an access unit delimiter, an end of sequence NAL unit, an end of bitstream NAL unit, or a filler data NAL unit. Parameter sets may be needed for the reconstruction of decoded pictures, whereas many of the other non-VCL NAL units are not necessary for the reconstruction of decoded sample values.

[01 13] Parameters that remain unchanged through a coded video sequence may be included in a sequence parameter set. In addition to the parameters that may be needed by the decoding process, the sequence parameter set may optionally contain video usability information (VUI), which includes parameters that may be important for buffering, picture output timing, rendering, and resource reservation. In HE VC a sequence parameter set RBSP includes parameters that can be referred to by one or more picture parameter set RBSPs or one or more SEI NAL units containing a buffering period SEI message. A picture parameter set contains such parameters that are likely to be unchanged in several coded pictures. A picture parameter set RBSP may include parameters that can be referred to by the coded slice NAL units of one or more coded pictures.

[01 14] In HEVC, a video parameter set (VPS) may be defined as a syntax structure containing syntax elements that apply to zero or more entire coded video sequences as determined by the content of a syntax element found in the SPS referred to by a syntax element found in the PPS referred to by a syntax element found in each slice segment header.

[01 15] A video parameter set RBSP may include parameters that can be referred to by one or more sequence parameter set RBSPs.

[0116] The relationship and hierarchy between video parameter set (VPS), sequence parameter set (SPS), and picture parameter set (PPS) may be described as follows. VPS resides one level above SPS in the parameter set hierarchy and in the context of scalability and/or 3D video. VPS may include parameters that are common for all slices across all (scalability or view) layers in the entire coded video sequence. SPS includes the parameters that are common for all slices in a particular (scalability or view) layer in the entire coded video sequence, and may be shared by multiple (scalability or view) layers. PPS includes the parameters that are common for all slices in a particular layer representation (the representation of one scalability or view layer in one access unit) and are likely to be shared by all slices in multiple layer representations.

[0117] VPS may provide information about the dependency relationships of the layers in a bitstream, as well as many other information that are applicable to all slices across all (scalability or view) layers in the entire coded video sequence. VPS may be considered to comprise two parts, the base VPS and a VPS extension, where the VPS extension may be optionally present.

[0118] Out-of-band transmission, signaling or storage can additionally or alternatively be used for other purposes than tolerance against transmission errors, such as ease of access or session negotiation. For example, a sample entry of a track in a file conforming to the ISO Base Media File Format may comprise parameter sets, while the coded data in the bitstream is stored elsewhere in the file or in another file. The phrase along the bitstream (e.g. indicating along the bitstream) or along a coded unit of a bitstream (e.g. indicating along a coded tile) may be used in claims and described embodiments to refer to out-of-band transmission, signaling, or storage in a manner that the out-of-band data is associated with the bitstream or the coded unit, respectively. The phrase decoding along the bitstream or along a coded unit of a bitstream or alike may refer to decoding the referred out-of-band data (which may be obtained from out-of-band transmission, signaling, or storage) that is associated with the bitstream or the coded unit, respectively. For example, the phrase along the bitstream may be used when the bitstream is contained in a file, such as a file conforming to the ISO Base Media File Format, and certain file metadata is stored in the file in a manner that associates the metadata to the bitstream, such as boxes in the sample entry for a track containing the bitstream, a sample group for the track containing the bitstream, or a timed metadata track associated with the track containing the bitstream.

[01 19] A SEI NAL unit may contain one or more SEI messages, which are not required for the decoding of output pictures but may assist in related processes, such as picture output timing, rendering, error detection, error concealment, and resource reservation. Several SEI messages are specified in H.264/AVC and HEVC, and the user data SEI messages enable organizations and companies to specify SEI messages for their own use. H.264/AVC and HEVC contain the syntax and semantics for the specified SEI messages but no process for handling the messages in the recipient is defined. Consequently, encoders are required to follow the H.264/AVC standard or the HEVC standard when they create SEI messages, and decoders conforming to the H.264/AVC standard or the HEVC standard, respectively, are not required to process SEI messages for output order conformance. One of the reasons to include the syntax and semantics of SEI messages in H.264/AVC and HEVC is to allow different system specifications to interpret the supplemental information identically and hence interoperate. It is intended that system specifications can require the use of particular SEI messages both in the encoding end and in the decoding end, and additionally the process for handling particular SEI messages in the recipient can be specified.

[0120] In HEVC, there are two types of SEI NAL units, namely the suffix SEI NAL unit and the prefix SEI NAL unit, having a different nal unit type value from each other. The SEI message(s) contained in a suffix SEI NAL unit are associated with the VCL NAL unit preceding, in decoding order, the suffix SEI NAL unit. The SEI message(s) contained in a prefix SEI NAL unit are associated with the VCL NAL unit following, in decoding order, the prefix SEI NAL unit.

[0121] A coded picture is a coded representation of a picture.

[0122] In HEVC, a coded picture may be defined as a coded representation of a picture containing all coding tree units of the picture. In HEVC, an access unit (AU) may be defined as a set of NAL units that are associated with each other according to a specified classification rule, are consecutive in decoding order, and contain at most one picture with any specific value of nuh layer id. In addition to containing the VCL NAL units of the coded picture, an access unit may also contain non- VCL NAL units. Said specified classification rule may for example associate pictures with the same output time or picture output count value into the same access unit.

[0123] A bitstream may be defined as a sequence of bits, in the form of a NAL unit stream or a byte stream, that forms the representation of coded pictures and associated data forming one or more coded video sequences. A first bitstream may be followed by a second bitstream in the same logical channel, such as in the same file or in the same connection of a

communication protocol. An elementary stream (in the context of video coding) may be defined as a sequence of one or more bitstreams. The end of the first bitstream may be indicated by a specific NAL unit, which may be referred to as the end of bitstream (EOB) NAL unit and which is the last NAL unit of the bitstream. In HEVC and its current draft extensions, the EOB NAL unit is required to have nuh layer id equal to 0.

[0124] In H.264/AVC, a coded video sequence is defined to be a sequence of consecutive access units in decoding order from an IDR access unit, inclusive, to the next IDR access unit, exclusive, or to the end of the bitstream, whichever appears earlier.

[0125] In HEVC, a coded video sequence (CVS) may be defined, for example, as a sequence of access units that consists, in decoding order, of an IRAP access unit with NoRaslOutputFlag equal to 1 , followed by zero or more access units that are not IRAP access units with NoRaslOutputFlag equal to 1 , including all subsequent access units up to but not including any subsequent access unit that is an IRAP access unit with NoRaslOutputFlag equal to 1. An IRAP access unit may be defined as an access unit in which the base layer picture is an IRAP picture. The value of NoRaslOutputFlag is equal to 1 for each IDR picture, each BLA picture, and each IRAP picture that is the first picture in that particular layer in the bitstream in decoding order, is the first IRAP picture that follows an end of sequence NAL unit having the same value of nuh layer id in decoding order. There may be means to provide the value of HandleCraAsBlaFlag to the decoder from an external entity, such as a player or a receiver, which may control the decoder. HandleCraAsBlaFlag may be set to 1 for example by a player that seeks to a new position in a bitstream or tunes into a broadcast and starts decoding and then starts decoding from a CRA picture. When HandleCraAsBlaFlag is equal to 1 for a CRA picture, the CRA picture is handled and decoded as if it were a BLA picture.

[0126] In HEVC, a coded video sequence may additionally or alternatively (to the specification above) be specified to end, when a specific NAL unit, which may be referred to as an end of sequence (EOS) NAL unit, appears in the bitstream and has nuh layer id equal to 0.

[0127] A group of pictures (GOP) and its characteristics may be defined as follows. A GOP can be decoded regardless of whether any previous pictures were decoded. An open GOP is such a group of pictures in which pictures preceding the initial intra picture in output order might not be correctly decodable when the decoding starts from the initial intra picture of the open GOP. In other words, pictures of an open GOP may refer (in inter prediction) to pictures belonging to a previous GOP. An HEVC decoder can recognize an intra picture starting an open GOP, because a specific NAL unit type, CRA NAL unit type, may be used for its coded slices. A closed GOP is such a group of pictures in which all pictures can be correctly decoded when the decoding starts from the initial intra picture of the closed GOP. In other words, no picture in a closed GOP refers to any pictures in previous GOPs. In

H.264/AVC and HEVC, a closed GOP may start from an IDR picture. In HEVC a closed GOP may also start from a BLA W RADL or a BLA N LP picture. An open GOP coding structure is potentially more efficient in the compression compared to a closed GOP coding structure, due to a larger flexibility in selection of reference pictures.

[0128] A Decoded Picture Buffer (DPB) may be used in the encoder and/or in the decoder. There are two reasons to buffer decoded pictures, for references in inter prediction and for reordering decoded pictures into output order. As H.264/AVC and HEVC provide a great deal of flexibility for both reference picture marking and output reordering, separate buffers for reference picture buffering and output picture buffering may waste memory resources. Hence, the DPB may include a unified decoded picture buffering process for reference pictures and output reordering. A decoded picture may be removed from the DPB when it is no longer used as a reference and is not needed for output.

[0129] In many coding modes of H.264/AVC and HEVC, the reference picture for inter prediction is indicated with an index to a reference picture list. The index may be coded with variable length coding, which usually causes a smaller index to have a shorter value for the corresponding syntax element. In H.264/AVC and HEVC, two reference picture lists

(reference picture list 0 and reference picture list 1) are generated for each bi-predictive (B) slice, and one reference picture list (reference picture list 0) is formed for each inter-coded (P) slice.

[0130] Many coding standards, including H.264/AVC and HEVC, may have decoding process to derive a reference picture index to a reference picture list, which may be used to indicate which one of the multiple reference pictures is used for inter prediction for a particular block. A reference picture index may be coded by an encoder into the bitstream is some inter coding modes or it may be derived (by an encoder and a decoder) for example using neighboring blocks in some other inter coding modes.

[0131] Several candidate motion vectors may be derived for a single prediction unit. For example, motion vector prediction HEVC includes two motion vector prediction schemes, namely the advanced motion vector prediction (AMVP) and the merge mode. In the AMVP or the merge mode, a list of motion vector candidates is derived for a PU. There are two kinds of candidates: spatial candidates and temporal candidates, where temporal candidates may also be referred to as TMVP candidates.

[0132] A candidate list derivation may be performed for example as follows, while it should be understood that other possibilities may exist for candidate list derivation. If the occupancy of the candidate list is not at maximum, the spatial candidates are included in the candidate list first if they are available and not already exist in the candidate list. After that, if occupancy of the candidate list is not yet at maximum, a temporal candidate is included in the candidate list. If the number of candidates still does not reach the maximum allowed number, the combined bi-predictive candidates (for B slices) and a zero motion vector are added in. After the candidate list has been constructed, the encoder decides the final motion information from candidates for example based on a rate-distortion optimization (RDO) decision and encodes the index of the selected candidate into the bitstream. Likewise, the decoder decodes the index of the selected candidate from the bitstream, constructs the candidate list, and uses the decoded index to select a motion vector predictor from the candidate list.

[0133] In HEVC, AMVP and the merge mode may be characterized as follows. In AMVP, the encoder indicates whether uni-prediction or bi-prediction is used and which reference pictures are used as well as encodes a motion vector difference. In the merge mode, only the chosen candidate from the candidate list is encoded into the bitstream indicating the current prediction unit has the same motion information as that of the indicated predictor. Thus, the merge mode creates regions composed of neighbouring prediction blocks sharing identical motion information, which is only signalled once for each region.

[0134] An example of the operation of advanced motion vector prediction is provided in the following, while other similar realizations of advanced motion vector prediction are also possible for example with different candidate position sets and candidate locations with candidate position sets. It also needs to be understood that other prediction mode, such as the merge mode, may operate similarly. Two spatial motion vector predictors (MVPs) may be derived and a temporal motion vector predictor (TMVP) may be derived. They may be selected among the positions: three spatial motion vector predictor candidate positions located above the current prediction block (Bo, Bi, B₂) and two on the left (Ao, Ai). The first motion vector predictor that is available (e.g. resides in the same slice, is inter-coded, etc.) in a pre defined order of each candidate position set, (Bo, Bi, B₂) or (Ao, Ai), may be selected to represent that prediction direction (up or left) in the motion vector competition. A reference index for the temporal motion vector predictor may be indicated by the encoder in the slice header (e.g. as a collocated ref idx syntax element). The first motion vector predictor that is available (e.g. is inter-coded) in a pre-defined order of potential temporal candidate locations, e.g. in the order (Co, Ci), may be selected as a source for a temporal motion vector predictor. The motion vector obtained from the first available candidate location in the co-located picture may be scaled according to the proportions of the picture order count differences of the reference picture of the temporal motion vector predictor, the co-located picture, and the current picture. Moreover, a redundancy check may be performed among the candidates to remove identical candidates, which can lead to the inclusion of a zero motion vector in the candidate list. The motion vector predictor may be indicated in the bitstream for example by indicating the direction of the spatial motion vector predictor (up or left) or the selection of the temporal motion vector predictor candidate. The co-located picture may also be referred to as the collocated picture, the source for motion vector prediction, or the source picture for motion vector prediction. [0135] Motion parameter types or motion information may include but are not limited to one or more of the following types:

[0136] an indication of a prediction type (e.g. intra prediction, uni-prediction, bi- prediction) and/or a number of reference pictures;

[0137] an indication of a prediction direction, such as inter (a.k.a. temporal) prediction, inter-layer prediction, inter-view prediction, view synthesis prediction (VSP), and inter component prediction (which may be indicated per reference picture and/or per prediction type and where in some embodiments inter-view and view-synthesis prediction may be jointly considered as one prediction direction) and/or

[0138] an indication of a reference picture type, such as a short-term reference picture and/or a long-term reference picture and/or an inter-layer reference picture (which may be indicated e.g. per reference picture)

[0139] a reference index to a reference picture list and/or any other identifier of a reference picture (which may be indicated e.g. per reference picture and the type of which may depend on the prediction direction and/or the reference picture type and which may be accompanied by other relevant pieces of information, such as the reference picture list or alike to which reference index applies);

[0140] a horizontal motion vector component (which may be indicated e.g. per prediction block or per reference index or alike);

[0141] a vertical motion vector component (which may be indicated e.g. per prediction block or per reference index or alike);

[0142] one or more parameters, such as picture order count difference and/or a relative camera separation between the picture containing or associated with the motion parameters and its reference picture, which may be used for scaling of the horizontal motion vector component and/or the vertical motion vector component in one or more motion vector prediction processes (where said one or more parameters may be indicated e.g. per each reference picture or each reference index or alike);

[0143] coordinates of a block to which the motion parameters and/or motion information applies, e.g. coordinates of the top-left sample of the block in luma sample units;

[0144] extents (e.g. a width and a height) of a block to which the motion parameters and/or motion information applies.

[0145] In general, motion vector prediction mechanisms, such as those motion vector prediction mechanisms presented above as examples, may include prediction or inheritance of certain pre-defined or indicated motion parameters. [0146] A motion field associated with a picture may be considered to comprise of a set of motion information produced for every coded block of the picture. A motion field may be accessible by coordinates of a block, for example. A motion field may be used for example in TMVP or any other motion prediction mechanism where a source or a reference for prediction other than the current (de)coded picture is used.

[0147] Different spatial granularity or units may be applied to represent and/or store a motion field. For example, a regular grid of spatial units may be used. For example, a picture may be divided into rectangular blocks of certain size (with the possible exception of blocks at the edges of the picture, such as on the right edge and the bottom edge). For example, the size of the spatial unit may be equal to the smallest size for which a distinct motion can be indicated by the encoder in the bitstream, such as a 4x4 block in luma sample units. For example, a so-called compressed motion field may be used, where the spatial unit may be equal to a pre-defined or indicated size, such as a 16x16 block in luma sample units, which size may be greater than the smallest size for indicating distinct motion. For example, an HEVC encoder and/or decoder may be implemented in a manner that a motion data storage reduction (MDSR) or motion field compression is performed for each decoded motion field (prior to using the motion field for any prediction between pictures). In an HEVC

implementation, MDSR may reduce the granularity of motion data to 16x16 blocks in luma sample units by keeping the motion applicable to the top-left sample of the 16x16 block in the compressed motion field. The encoder may encode indication(s) related to the spatial unit of the compressed motion field as one or more syntax elements and/or syntax element values for example in a sequence-level syntax structure, such as a video parameter set or a sequence parameter set. In some (de)coding methods and/or devices, a motion field may be represented and/or stored according to the block partitioning of the motion prediction (e.g. according to prediction units of the HEVC standard). In some (de)coding methods and/or devices, a combination of a regular grid and block partitioning may be applied so that motion associated with partitions greater than a pre-defined or indicated spatial unit size is represented and/or stored associated with those partitions, whereas motion associated with partitions smaller than or unaligned with a pre-defined or indicated spatial unit size or grid is represented and/or stored for the pre-defined or indicated units.

[0148] Scalable video coding may refer to coding structure where one bitstream can contain multiple representations of the content, for example, at different bitrates, resolutions or frame rates. In these cases the receiver can extract the desired representation depending on its characteristics (e.g. resolution that matches best the display device). Alternatively, a server or a network element can extract the portions of the bitstream to be transmitted to the receiver depending on e.g. the network characteristics or processing capabilities of the receiver. A meaningful decoded representation can be produced by decoding only certain parts of a scalable bit stream. A scalable bitstream typically consists of a“base layer” providing the lowest quality video available and one or more enhancement layers that enhance the video quality when received and decoded together with the lower layers. In order to improve coding efficiency for the enhancement layers, the coded representation of that layer typically depends on the lower layers. E.g. the motion and mode information of the enhancement layer can be predicted from lower layers. Similarly the pixel data of the lower layers can be used to create prediction for the enhancement layer.

[0149] In some scalable video coding schemes, a video signal can be encoded into a base layer and one or more enhancement layers. An enhancement layer may enhance, for example, the temporal resolution (i.e., the frame rate), the spatial resolution, or simply the quality of the video content represented by another layer or part thereof. Each layer together with all its dependent layers is one representation of the video signal, for example, at a certain spatial resolution, temporal resolution and quality level. In this document, we refer to a scalable layer together with all of its dependent layers as a“scalable layer representation”. The portion of a scalable bitstream corresponding to a scalable layer representation can be extracted and decoded to produce a representation of the original signal at certain fidelity.

[0150] Scalability modes or scalability dimensions may include but are not limited to the following:

[0151] Quality scalability: Base layer pictures are coded at a lower quality than

enhancement layer pictures, which may be achieved for example using a greater quantization parameter value (i.e., a greater quantization step size for transform coefficient quantization) in the base layer than in the enhancement layer. Quality scalability may be further categorized into fine-grain or fine-granularity scalability (FGS), medium-grain or medium-granularity scalability (MGS), and/or coarse-grain or coarse-granularity scalability (CGS), as described below.

[0152] Spatial scalability: Base layer pictures are coded at a lower resolution (i.e. have fewer samples) than enhancement layer pictures. Spatial scalability and quality scalability, particularly its coarse-grain scalability type, may sometimes be considered the same type of scalability.

[0153] Bit-depth scalability: Base layer pictures are coded at lower bit-depth (e.g. 8 bits) than enhancement layer pictures (e.g. 10 or 12 bits). [0154] Dynamic range scalability: Scalable layers represent a different dynamic range and/or images obtained using a different tone mapping function and/or a different optical transfer function.

[0155] Chroma format scalability: Base layer pictures provide lower spatial resolution in chroma sample arrays (e.g. coded in 4:2:0 chroma format) than enhancement layer pictures (e.g. 4:4:4 format).

[0156] Color gamut scalability: enhancement layer pictures have a richer/broader color representation range than that of the base layer pictures - for example the enhancement layer may have UHDTV (ITU-R BT.2020) color gamut and the base layer may have the ITU-R BT.709 color gamut.

[0157] View scalability, which may also be referred to as multiview coding. The base layer represents a first view, whereas an enhancement layer represents a second view. A view may be defined as a sequence of pictures representing one camera or viewpoint. It may be considered that in stereoscopic or two-view video, one video sequence or view is presented for the left eye while a parallel view is presented for the right eye.

[0158] Depth scalability, which may also be referred to as depth-enhanced coding. A layer or some layers of a bitstream may represent texture view(s), while other layer or layers may represent depth view(s).

[0159] Region-of- interest scalability (as described below).

[0160] Interlaced-to-progressive scalability (also known as field-to-ffame scalability): coded interlaced source content material of the base layer is enhanced with an enhancement layer to represent progressive source content. The coded interlaced source content in the base layer may comprise coded fields, coded frames representing field pairs, or a mixture of them. In the interlace-to-progressive scalability, the base-layer picture may be resampled so that it becomes a suitable reference picture for one or more enhancement- layer pictures.

[0161] Hybrid codec scalability (also known as coding standard scalability): In hybrid codec scalability, the bitstream syntax, semantics and decoding process of the base layer and the enhancement layer are specified in different video coding standards. Thus, base layer pictures are coded according to a different coding standard or format than enhancement layer pictures. For example, the base layer may be coded with H.264/AVC and an enhancement layer may be coded with an HEVC multi-layer extension.

[0162] It should be understood that many of the scalability types may be combined and applied together. For example color gamut scalability and bit-depth scalability may be combined. [0163] The term layer may be used in context of any type of scalability, including view scalability and depth enhancements. An enhancement layer may refer to any type of an enhancement, such as SNR, spatial, multiview, depth, bit-depth, chroma format, and/or color gamut enhancement. A base layer may refer to any type of a base video sequence, such as a base view, a base layer for SNR/spatial scalability, or a texture base view for depth-enhanced video coding.

[0164] Some scalable video coding schemes may require IRAP pictures to be aligned across layers in a manner that either all pictures in an access unit are IRAP pictures or no picture in an access unit is an IRAP picture. Other scalable video coding schemes, such as the multi-layer extensions of HEVC, may allow IRAP pictures that are not aligned, i.e. that one or more pictures in an access unit are IRAP pictures, while one or more other pictures in an access unit are not IRAP pictures. Scalable bitstreams with IRAP pictures or similar that are not aligned across layers may be used for example for providing more frequent IRAP pictures in the base layer, where they may have a smaller coded size due to e.g. a smaller spatial resolution. A process or mechanism for layer- wise start-up of the decoding may be included in a video decoding scheme. Decoders may hence start decoding of a bitstream when a base layer contains an IRAP picture and step-wise start decoding other layers when they contain IRAP pictures. In other words, in a layer-wise start-up of the decoding mechanism or process, decoders progressively increase the number of decoded layers (where layers may represent an enhancement in spatial resolution, quality level, views, additional components such as depth, or a combination) as subsequent pictures from additional enhancement layers are decoded in the decoding process. The progressive increase of the number of decoded layers may be perceived for example as a progressive improvement of picture quality (in case of quality and spatial scalability).

[0165] A sender, a gateway, a client, or another entity may select the transmitted layers and/or sub-layers of a scalable video bitstream. Terms layer extraction, extraction of layers, or layer down-switching may refer to transmitting fewer layers than what is available in the bitstream received by the sender, the gateway, the client, or another entity. Layer up- switching may refer to transmitting additional layer(s) compared to those transmitted prior to the layer up-switching by the sender, the gateway, the client, or another entity, i.e. restarting the transmission of one or more layers whose transmission was ceased earlier in layer down switching. Similarly to layer down-switching and/or up-switching, the sender, the gateway, the client, or another entity may perform down- and/or up-switching of temporal sub-layers. The sender, the gateway, the client, or another entity may also perform both layer and sub- layer down-switching and/or up-switching. Layer and sub-layer down-switching and/or up- switching may be carried out in the same access unit or alike (i.e. virtually simultaneously) or may be carried out in different access units or alike (i.e. virtually at distinct times).

[0166] Scalability may be enabled in two basic ways. Either by introducing new coding modes for performing prediction of pixel values or syntax from lower layers of the scalable representation or by placing the lower layer pictures to a reference picture buffer (e.g. a decoded picture buffer, DPB) of the higher layer. The first approach may be more flexible and thus may provide better coding efficiency in most cases. However, the second, reference frame based scalability, approach may be implemented efficiently with minimal changes to single layer codecs while still achieving majority of the coding efficiency gains available. Essentially a reference frame based scalability codec may be implemented by utilizing the same hardware or software implementation for all the layers, just taking care of the DPB management by external means.

[0167] A scalable video encoder for quality scalability (also known as Signal-to-Noise or SNR) and/or spatial scalability may be implemented as follows. For a base layer, a conventional non-scalable video encoder and decoder may be used. The

reconstructed/decoded pictures of the base layer are included in the reference picture buffer and/or reference picture lists for an enhancement layer. In case of spatial scalability, the reconstructed/decoded base-layer picture may be upsampled prior to its insertion into the reference picture lists for an enhancement-layer picture. The base layer decoded pictures may be inserted into a reference picture list(s) for coding/decoding of an enhancement layer picture similarly to the decoded reference pictures of the enhancement layer. Consequently, the encoder may choose a base-layer reference picture as an inter prediction reference and indicate its use with a reference picture index in the coded bitstream. The decoder decodes from the bitstream, for example from a reference picture index, that a base-layer picture is used as an inter prediction reference for the enhancement layer. When a decoded base-layer picture is used as the prediction reference for an enhancement layer, it is referred to as an inter- layer reference picture.

[0168] While the previous paragraph described a scalable video codec with two scalability layers with an enhancement layer and a base layer, it needs to be understood that the description can be generalized to any two layers in a scalability hierarchy with more than two layers. In this case, a second enhancement layer may depend on a first enhancement layer in encoding and/or decoding processes, and the first enhancement layer may therefore be regarded as the base layer for the encoding and/or decoding of the second enhancement layer. Furthermore, it needs to be understood that there may be inter-layer reference pictures from more than one layer in a reference picture buffer or reference picture lists of an enhancement layer, and each of these inter-layer reference pictures may be considered to reside in a base layer or a reference layer for the enhancement layer being encoded and/or decoded.

Furthermore, it needs to be understood that other types of inter-layer processing than reference- layer picture upsampling may take place instead or additionally. For example, the bit-depth of the samples of the reference- layer picture may be converted to the bit-depth of the enhancement layer and/or the sample values may undergo a mapping from the color space of the reference layer to the color space of the enhancement layer.

[0169] Inter- layer prediction may be defined as prediction in a manner that is dependent on data elements (e.g., sample values or motion vectors) of reference pictures from a different layer than the layer of the current picture (being encoded or decoded). Many types of inter layer prediction exist and may be applied in a scalable video encoder/decoder. The available types of inter-layer prediction may for example depend on the coding profile according to which the bitstream or a particular layer within the bitstream is being encoded or, when decoding, the coding profile that the bitstream or a particular layer within the bitstream is indicated to conform to. Alternatively or additionally, the available types of inter-layer prediction may depend on the types of scalability or the type of an scalable codec or video coding standard amendment (e.g. SHVC, MV-HEVC, or 3D-HEVC) being used.

[0170] A direct reference layer may be defined as a layer that may be used for inter- layer prediction of another layer for which the layer is the direct reference layer. A direct predicted layer may be defined as a layer for which another layer is a direct reference layer. An indirect reference layer may be defined as a layer that is not a direct reference layer of a second layer but is a direct reference layer of a third layer that is a direct reference layer or indirect reference layer of a direct reference layer of the second layer for which the layer is the indirect reference layer. An indirect predicted layer may be defined as a layer for which another layer is an indirect reference layer. An independent layer may be defined as a layer that does not have direct reference layers. In other words, an independent layer is not predicted using inter-layer prediction. A non-base layer may be defined as any other layer than the base layer, and the base layer may be defined as the lowest layer in the bitstream. An independent non-base layer may be defined as a layer that is both an independent layer and a non-base layer.

[0171] Recently, the development of various artificial neural network (NN) techniques, especially the ones related to deep learning, has enabled to leam algorithms for several tasks from the raw data, which algorithms may outperform algorithms which have been developed for many years using traditional (non-learning based) methods.

[0172] Artificial neural networks, or simply neural networks, are parametric computation graphs consisting of units and connections. The units may be arranged in successive layers, and in some neural network architectures only units in adjacent layers are connected. Each connection has an associated parameter or weight, which defines the strength of the connection. The weight gets multiplied by the incoming signal in that connection. In fully- connected layers of a feedforward neural network, each unit in a layer is connected to each unit in the following layer. So, the signal which is output by a certain unit gets multiplied by the connections connecting that unit to another unit in the following layer. The latter unit then may perform a simple operation such as a sum of the weighted signals.

[0173] Apart from fully-connected layers, there are different types of layers, such as but not limited to convolutional layers, non-linear activation layers, batch-normalization layers, and pooling layers.

[0174] The input layer receives the input data, such as images, and the output layer is task- specific and outputs an estimate of the desired data, for example a vector whose values represent a class distribution in the case of image classification. The“quality” of the neural network’s output is evaluated by comparing it to ground-truth output data. The comparison may include a loss or cost function, run on the neural network’s output and the ground-truth data. This comparison would then provide a“loss” or“cost” value.

[0175] The weights of the connections represent the biggest part of the leamable parameters of a neural network. Hereinafter, the terms“model“ and“neural network" are used interchangeably, as well as the weights of neural networks are sometimes referred to as leamable parameters or simply as parameters.

[0176] The parameters are learned by means of a training algorithm, where the goal is to minimize the loss value on a training dataset and on a held-out validation dataset. In order to minimize such value, the network is ran on a training dataset, a loss value is computed for the whole training dataset or for part of it, and the leamable parameters are modified in order to minimize the loss value on the training dataset. However, the performance of the training is evaluated on the held-out validation dataset. The training dataset is regarded as a

representative sample of the whole data. One popular learning approach is based on iterative local methods, where the loss on the training dataset is minimized by following the negative gradient direction. Here, the gradient is understood to be the gradient of the loss with respect to the leamable parameters of the neural network. The loss may be represented by the reconstructed prediction error. Computing the gradient on the whole training dataset may be computationally too heavy, thus learning is performed in sub-steps, where at each step a mini batch of data is sampled and gradients are computed from the mini-batch. This is referred to as stochastic gradient descent. The gradients are usually computed by back-propagation algorithm, where errors are propagated from the output layer to the input layer, by using the chain rule for differentiation. If the loss function or some operations performed by the neural network are not differentiable, it is still possible to estimate the gradient of the loss by using policy gradient methods, such as those used in reinforcement learning. The computed gradients are then used by one of the available optimization routines (such as stochastic gradient descent, Adam, RMSprop, etc.), to compute a weight update, which is then applied to update the weights of the network. After a full pass over the training dataset, the process is repeated several times until a convergence criterion is met, usually a generalization criterion. A generalization criterion may be derived from the loss value on the held-out validation dataset, for example by stopping the training when the loss value on the held-out validation dataset is less than a certain threshold. The gradients of the loss, i.e., the gradients of the reconstructed prediction error with respect to the weights of the neural network, may be referred to as the training signal.

[0177] Training a neural network is an optimization process, but as a difference to a typical optimization where the only goal is to minimize a function, the goal of the optimization or training process in machine learning is to make the model to learn the properties of the data distribution. In other words, the goal is to learn to generalize to previously unseen data, i.e., data which was not used for training the model. This is usually referred to as generalization.

In practice, data is usually split into two (or more) sets, the training set and the validation set. The training set is used for training the network, i.e., to modify its leamable parameters to minimize the loss. The validation set is used for checking the performance of the network on data which was not used to minimize the loss, as an indication of the final performance of the model. In particular, the errors on the training set and on the validation set are monitored during the training process to understand the following issues:

If the network is learning at all - in this case, the training set error should decrease, otherwise we are in the regime of underfitting.

If the network is learning to generalize - in this case, also the validation set error needs to decrease and to be not too much higher than the training set error. If the training set error is low, but the validation set error is much higher than the training set error, or it does not decrease, or it even increases, the model is in the regime of overfitting. This means that the model has just memorized the training set’s properties and performs well only on that set, but performs poorly on a set not used for tuning its parameters.

[0178] In general, overfitting is not a desirable property and is actually an issue to be avoided, as a model typically needs to perform well on previously unknown data as well. However, in certain applications where the neural networks are fine-tuned for a specific type of data, overfitting may provide superior results.

[0179] For example, neural networks based video processing may be used for enhancing a received video in the receiving device such that the device transmitting the video may send the video at lower quality. The transmitting device may comprise a video encoder and the receiving device may comprise a video decoder. The transmitting device may degrade the quality of the video to be transmitted for example as follows:

- the resolution of the video may be downscaled and sent to the receiving device, which may use so-called super-resolution (SR) neural networks to upscale the lowered resolution of the video back to the original resolution;

the transmitting device may encode the video using a certain encoder or a set of encoder parameters, which introduces coding artifacts, and the receiving device post processes the received video to remove such artifacts. One of the sources of coding artifacts may be the quantization process;

the transmitting device may, for each video frame, remove areas which are easily inpaintable by a neural network present at the receiver’s side, and then encode and send only the non-removed areas of the video frames. Similarly, rather than removing areas that are easily inpaintable, the transmitting device may encode them in a manner that consumes a very small number of bits, e.g. using a prediction mode that can be signaled with very few bits (e.g., the merge mode in HEVC inter coding or a DC prediction mode in intra prediction) and omitting the prediction error coding. The receiving device then uses a neural network to inpaint the missing or specially coded areas to the received frames.

[0180] For all the above examples of processing for enhancing the video quality, neural networks may be trained to be especially effective as fine-tuned (i.e., further trained) on the actual target video on which they are used later on. In other words, the neural networks may be efficiently overfitted on the target content/video. However, overfitting may be performed in a relatively straightforward way only in the transmitting device, which has availability of the original data, which is used as ground-truth or target data during the training process. The overfitted neural network may then be sent to the receiving device to be used in enhancing the quality of the video.

[0181] In the neural networks described above, a large memory storage is typically needed for the weights of the neural network. Moreover, the number of operations to be performed is also very high. Thus, training neural networks is a demanding task in terms of the required computations, memory consumption and processing power, usually requiring specific hardware, such as GPUs, occupying relatively large physical space. Such capabilities may not always be available, especially in small devices such as IoT devices, surveillance cameras, mobile phones, etc. In other words, such capability-restricted device cannot overfit the neural networks on the target content/video, and as a result, the transmitting device cannot reduce the quality of the video (and thereby the bitrate), whereupon a wide communication bandwidth is required to transmit the video to the receiving device. Furthermore, even if the transmitting device had training capabilities, it may send the neural network to the receiver device. Neural network’s weights may require a large number of bits to be transmitted, thus the bandwidth between the transmitter and the receiver may be another limiting factor to the case where the transmitter fine-tuned the neural network on the target content.

[0182] An enhanced method for neural network-based enhancing of video content, especially if originated from bandwidth, memory and/or computation limited devices, is introduced herein.

[0183] An example of a method, which is depicted in the flow chart of Figure 5, comprises receiving (500), in a first apparatus, a first portion of a video data stream at a predetermined high quality; using (502) the high quality first portion of the video data stream to train a neural network for enhancing a video data stream of a type sufficiently similar to the first portion of the video data stream; receiving (504), in the first apparatus, a second portion of said video data stream at a predetermined lower quality; and enhancing (506) the quality of the lower quality second portion of said video data stream using the trained neural network. Streaming or progressive delivery may be used in various video services or applications, including but not limited to streaming services, point-to-point or multi-point video

conferencing, video surveillance or remote monitoring, and progressive downloading of video files (i.e., playing a file while downloading it).

[0184] Thus, the method addresses the problem of streaming or progressive delivery of a video in a system, where the capabilities of the transmitting device are limited in at least one of computation, power and memory capabilities, and the communication bandwidth between the transmitting and the receiving device is limited too. The receiving device may preferably have less limited capabilities so as to be capable of training neural networks.

[0185] The transmitting device may initially send the video frames at high quality, where high quality may refer to one or more predetermined parameter values defining, for example, high resolution, low quantization parameters, or no missing areas within each frame, etc.

[0186] The receiving device uses the high quality frames, besides for the playback of the video frames, also for training a neural network to enhance a similar type video data stream. For this task, the high quality frames may be used as ground-truth for overfitting the neural network, and the input to the neural network may be a down-graded version of the received high-quality frames, where the down-grading operation may include encoding and decoding the data with lower quality, removing areas from the image, or reducing the resolution. In one embodiment, one neural network for each of these different quality modification techniques may be trained, and each such neural network would be associated with a unique ID which may be agreed with or anyway understood by the transmitting device. During training, the obtained loss for each of these neural networks is recorded so that it is possible to later understand which of these neural networks performs best on the content used for training.

[0187] Subsequently, the transmitting device starts to send the video frames at a lower quality, whereupon the receiving device uses the trained neural network for enhancing the received lower quality video frames back to high quality frames. In one embodiment, if multiple neural networks were trained, where each neural network was trained to enhance a frame which was down-graded using a different technique, the receiver device first signals to the transmitter device to use the down-grading technique for which training was most successful, and then the receiver device uses the corresponding neural network for enhancing the received low quality content.

[0188] Figure 6 illustrates a simplified example of a setup for implementing the method and the embodiments related thereto. In this example, a transmitting device 600, also referred to as the second apparatus, a sender or Device A, is capable of capturing video, but may be limited in communication bandwidth at least in respect to a receiving device 602, also referred to as the first apparatus, a receiver or Device B. The transmitting device may also be limited in terms of one or more of the following capabilities: computation capability, available memory and processing power. For example, the transmitting device may lack such a graphics processing unit (GPU) that would be required for training a neural network in real-time. The term“limited” in this context may especially refer to the required capabilities for training neural networks, and/or limited by other factors such as the need to preserve power consumption and limited in available physical space for computation, memory and power related hardware. In some embodiments, in addition to or instead of being capable of capturing video, Device A is capable of generating coded video bitstream through other means, such as transcoding an incoming video bitstream or encoding screen content, such as computer game content or a multimedia presentation, as a video bitstream.

[0189] It is noted that the above limitations are described only to emphasize the advantages obtained by the embodiments in certain device setups. Nevertheless, the embodiments are equally applicable in devices having no such limitations.

[0190] The transmitting device may, for example, comprise:

A surveillance or IoT camera, which may be battery-powered;

An encoding device operationally connected to several cameras, such as surveillance or IoT cameras, wherein the input from a subset (one or more) of cameras is selected for encoding at any particular time. The selection may be pre-configured (e.g. time- sliced round-robin fashion) or user-controlled;

A drone with a camera. In such a scenario, the battery consumption used for radio transmission may depend on the video bitrate, hence minimizing the bitrate is beneficial.

[0191] The transmitting device 600 may stream the captured video or transmit a file of the captured video to the receiving device 602. The transmitting device 600 is able to control the quality of the streamed video, in one or multiple ways, such as adjusting resolution, frame rate, quantization parameter, or removing pixel-areas from one or more frames. The receiving device 602 receives the video and performs one of the tasks for which it was designed for, such as video playback. The receiving device 602 does not preferably have the limitations of the transmitting device 600, in the sense that the receiving device 602 is able to perform training of neural network(s).

[0192] In the following, some embodiments relating to the operation of the receiving device 602 are described more in detail.

[0193] According to an embodiment, the method further comprises pre-processing the received high quality first portion of the video data stream into a lower quality; inputting the pre-processed lower quality first portion of the video data stream in said neural network; computing a loss of the neural network on the basis of a difference between a video data stream at an output of the neural network and the received high quality first portion of the video data stream; and deriving a training signal for the neural network based on the loss. [0194] Thus, for each high quality portion of the video data streamed by transmitting device 600, the receiving device 602 uses spatiotemporal unit(s), such as ffame(s), slice(s), or block(s) of the high quality video data for overfitting a neural network. Embodiments are generally not limited to any particular high quality portion or low quality portion. A portion of the video data may correspond to a set of spatiotemporal units. In different embodiments, a portion of the video data is one of the following or a combination thereof:

One or more pictures (which are typically frames but could likewise be fields)

One or more slices, or other spatial subsets of a picture

One or more blocks, such as coding units of HEVC or alike

[0195] For carrying out the overfitting, the received and decoded high quality

spatiotemporal unit(s) are pre-processed into a lower quality. This may comprise, for example, decreasing the resolution or the frame rate, or the frames may be re-encoded using a coarser quantization (i.e., higher quantization parameter), or pixel areas may be removed from each frame. The pre-processed spatiotemporal unit(s) are input to the neural network to be processed, whereas the originally received high quality spatiotemporal unit(s) are used as the ground truth in the output of the neural network for computing the loss. The loss may be computed, for example, as the mean squared error derived from difference of the ground truth and the output of the neural network. From the loss, a training signal may be derived, for example by using backpropagation for computing the gradient of the loss with respect to the neural network’s leamable parameters, and stochastic gradient descent to update the neural net’s leamable parameters.

[0196] Figure 7 depicts a video/image decoder 700 according to an aspect of the invention related to applying coarser quantization as means for pre-processing into a lower quality. Figure 7 describes the decoding of a spatiotemporal unit that is used as input for overfitting the neural network. Without loss of generality, the description refers to a block as the spatiotemporal unit, but likewise applies to other spatiotemporal units, such as a frame or a slice.

[0197] In an embodiment depicted in Figure 7, the coding parameters for an encoded version of the prediction error are entropy-decoded 702 from the bitstream. For example, the quantized transform coefficients 704 may be entropy-decoded from the bitstream.

[0198] A reconstructed prediction error is derived in the inverse quantization and transform unit 706, which may more generally be regard as prediction error decoding. For example, the quantized transform coefficients may be inverse-quantized and inverse- transformed in the inverse quantization and transform unit 706 to form a reconstructed prediction error block 708.

[0199] In the compensation process 710, the reconstructed prediction error block 708 is sample-wise summed up with the prediction block 730 to form a reconstructed reference (block) 712. The compensation process 710 may also comprise additional operations, such as filtering, e.g. deblocking loop filtering and/or sample adaptive offset filtering. The reconstructed reference block 712 may be stored as a part of a reconstructed reference picture to be used as input for subsequent predictions.

[0200] A reconstructed reference block 712 may be also used as the output of the decoder, or may be further processed e.g. by filtering (e.g. deblocking and/or sample adaptive offset filtering) and/or cropping.

[0201] In the second quantization process 714, the quantized transform coefficients are quantized more coarsely. A certain target quantization parameter value may be selected for this second quantization step 714.

[0202] A reconstructed prediction error 718 for the more coarsely quantized prediction error is derived. For example, the quantized transform coefficients may be inverse-quantized (compensating both the first and second quantization) and inverse-transformed a second inverse quantization and transform unit 716 to form a reconstructed prediction error block. The inverse-quantization may be performed in one or two steps. In the latter case, the inverse quantization is first performed as an inverse operation to the second (i.e. coarser)

quantization, and then inverse quantization is performed as an inverse operation to the first (regular) quantization that was performed in encoding.

[0203] The reconstructed prediction error for the more coarsely quantized prediction error is enhanced using the neural net 724. The output of this step is referred to as enhanced reconstructed prediction error 732.

[0204] A training signal is computed in a computing unit 720 by taking as inputs the reconstructed prediction error 708 and the enhanced reconstructed prediction error 732 In an embodiment, the training signal comprises the gradient of the loss, where the loss is represented by the difference of the reconstructed prediction error and the enhanced reconstructed prediction error. The training signal is used to fine-tune the neural net in a fine- tuning unit 722.

[0205] The predictor 726 in the decoder obtains prediction mode selections 728 from the entropy decoding 702 (and typically does not comprise a search for prediction mode parameters, like the encoder-side predictor might have). Otherwise, the predictor 726 produces prediction blocks 730 on the basis of the reconstructed reference signal 712. The predictor realizes a prediction process of a certain type, such as intra prediction and/or inter prediction, or modes thereof, according to the prediction mode selections 728.

[0206] Figure 8 depicts a video/image decoding for the low quality portion corresponding to Figure 8 for video/image decoding the high quality portion. Without loss of generality, the description refers to a block as the spatiotemporal unit, but likewise applies to other spatiotemporal units, such as a frame or a slice. The compensation and supplementary compensation steps operate otherwise identically but the compensation step takes the

(conventionally) reconstructed prediction error as input and the supplementary compensation step takes the enhanced reconstructed prediction error as input. The enhanced reconstructed reference that is output by the supplementary compensation may be used for output, such as displaying. Otherwise, the different steps operate similarly to what has been described earlier on respective steps.

[0207] According to an embodiment, the method further comprises training the neural network iteratively with an iteratively updated training signal until the loss is minimized. Due to the nature of neural networks, a certain number of training iterations may be performed to minimize the loss.

[0208] According to an embodiment, the loss is minimized when one or more

predetermined stopping criteria are met. Thus, the overfitting is performed by minimizing the loss until a certain stopping criterion is achieved. The stopping criterion may be, for example, a maximum number of training iterations, or it may be based on the difference in the loss value within a certain time budget. In the latter option, if the loss has not decreased sufficiently during a certain number of iterations, the training will be stopped. As the goal is to overfit on the training data (i.e. the high quality first portion of the video data stream), there is no need to monitor the error on a separate validation data set. However, monitoring such validation error may give an indication of how much the neural network has been overfitted. The amount of overfitting may be proportional to the gap between training error and validation error.

[0209] According to an embodiment, in response to the received high quality first portion of the video data stream being insufficient, e.g. temporally too short, for completing the training of the neural network, the method further comprises using a subsequent high quality portion of the video data stream for training the neural network. Hence, it may happen that a certain spatiotemporal unit which is streamed at high quality is insufficient, e.g. temporally too short, for successfully overfitting a neural network. For example, while overfitting the neural network on the incoming high quality frames, the receiver may monitor the amount of overfitting or just the training error. When a new temporal portion of the video data stream is received, the receiver will have to evaluate if the overfitting of the neural network for the previous temporal portion has been completed. If the amount of overfitting did not reach a predetermined threshold or the training error was not lower than a predetermined threshold, the receiver considers the neural network’s state as not yet completely overfitted, thus it will not be used at inference stage yet (i.e., it will not be used for enhancing similar shots). If the received new temporal portion of the video data stream is similar enough to the previous, such as the first, temporal portion, the receiver continues the overfitting process for that type of video data.

[0210] According to an embodiment, the method further comprises analyzing, in the first apparatus, the received video data stream; and dividing the received video data stream into the first and any subsequent portions. Upon receiving the bitstream, the receiving device decodes the frames and carries out video playback of the decoded frames. At the same time or alternatively before initiating playback of the received frames,, the receiving device analyses the received frames, either in the decoded domain or in the bitstream domain, in order to identify significant changes in the content of the video frames.

[0211 ] According to an embodiment, each high quality portion will be associated with its own overfitted neural network.

[0212] According to an embodiment, the method further comprises identifying scene boundaries in the received video data stream; and dividing the received video data stream temporally into the first and any subsequent portions at locations of identified scene boundaries. An example of significant changes in the content of the video frames is a change of a scene (a.k.a. a shot). Thus, the receiving device is configured to identify scene boundaries in the received video data stream. Such operation is usually referred to as shot or scene boundary detection, where an image similarity measure is used to compare frames and a shot boundary is detected for example when the similarity value between consecutive frames is below a threshold. Different methods may be applicable for detecting gradual scene transitions (fades, dissolves, etc.), such as comparing the similarity between non-consecutive frames separated by a duration proportional to the duration of typical gradual scene transitions. It is nevertheless noted that the embodiments are not limited to any specific shot boundary detection algorithm. As a result of the shot boundary detection, the receiving device divides the video into temporal portions, where a number of consecutive frames being similar in terms of the adopted similarity measure are determined to form a portion. [0213] After overfitting a neural net on a certain high quality temporal portion, said neural net may be used for processing, such as for enhancing, the quality of a subsequent temporal portion which is similar enough to the portion used for overfitting. Here, a realistic assumption is that in many types of videos there are recurring shots with very similar characteristics. Accordingly, the transmitting device does not need to send the subsequent temporal portions at high quality any more, but it may start to stream the subsequent temporal portions using lower quality. Here, by subsequent temporal portion we mean a portion which is similar to another portion which was previously transmitted to the receiver with high quality.

[0214] Figure 9 shows an illustration of how similar shots are streamed at different qualities. A shot may refer to a portion of video content having substantially similar visual properties, for example, a shot may be captured from the same camera at the same direction. A shot boundary may be therefore detected for example between portions captured with different cameras, or with the same camera at a different direction. Similar shots are denoted with the same number, e.g., all shots denoted as shot 1 are similar shots. Shots with different numbers are considered to be dissimilar. The first time when any of the shots 1, 2, 3, 4, 5 is sent, it is sent with high quality so that the receiving device can use the high quality shot for overfitting a neural network. The following times when any of the shots 1, 2, 3, 4, 5 is sent, it is sent with low quality and the receiving device will use the previously overfitted neural network for enhancing the low quality shot.

[0215] However, the transmitting device may be provided with information about when and for which temporal portion the lower quality streaming may be applied.

[0216] According to an embodiment, the receiving device signals information about the quality levels of the portions of the video stream to the transmitting device. This embodiment may be illustrated by referring to Figure 10. The transmitting device 1000 captures or otherwise acquires a video 1002 and controls the encoding 1004 of the video according to the quality control signals received from the receiving apparatus. The receiving apparatus 1006 receives the video stream and carries out the shot boundary detection 1010. For each new detected shot in the received video data stream, the receiving device analyses one or more decoded frames and determines 1012 if the shot is similar enough to one of the previous shots used for overfitting a neural network. This may be implemented by storing, for each high quality temporal portion used for training, one or more key frames as representatives of the shot, and comparing said key frames to the frames of the analyzed shot using a similarity metric. When a similar shot is detected, the receiving device selects 1014 a previously overfitted neural network for the shot and uses 1016 it for producing a neural network- enhanced shot. The receiving device also signals 1018 to the transmitting device that from now on the frames can be streamed using lower quality. If no previous similar shot is detected, the receiving device starts the overfitting process for the newly detected shot.

[0217] According to an embodiment, the receiver sends to the transmitter also information about which down-grading technique to use, such as lower resolution, higher quantization parameter, or removing part of the data, or a combination thereof. Also, the receiver may send additional information and parameters which are specific to the signaled down-grading technique, for example the down-sampling factor for the case of lowering the resolution, or the amount of data to be removed for the case of removing part of the data. The down-grading techniques may even be combined, for example sequentially. In this case, also the order of applying the down-grading techniques may be signaled from the receiver to the transmitter.

[0218] According to an embodiment, the quality levels for transmitting the video stream comprise a high quality level and one or more lower quality levels. Thus, in the simplest implementation, there may be only two quality levels, i.e. the high quality level and a low quality level, wherein the signaling may be carried out by one bit. On the other hand, multiple lower quality levels, such as multiple resolutions, may be easily implemented.

[0219] When detecting a subsequent shot, the receiving device again analyses one or more decoded frames and determines if the shot is similar enough to one of the previous shots used for overfitting a neural network. As a result, the receiving device either switches 1014 to use a different overfitted neural net if the new shot is similar enough to a shot previously used for overfitting said neural network, or the receiving device signals to the transmitting device that the frames of the current shot need to be streamed in higher quality.

[0220] According to an embodiment, the receiving device encodes, e.g. with a

conventional video or image encoder, and transmits one or more key frames or parts thereof as representatives of the shot to the transmitting device. A key frame may be selected for example from the low quality portion of a scene and thus a key frame has undergone neural net based enhancement prior to its encoding in the receiving device. A key frame may be selected e.g. periodically such that the transmitting device is able to check the progress and impact of neural net fine-tuning. The transmitting device selects the quality levels for the portions of the video stream based on the received key frames. In this embodiment, every time the receiving device overfits a neural network on a certain high quality temporal portion, the receiving device sends one or more key frames or parts thereof as representative of that portion to the transmitting device. Based on this, the transmitting device may conclude that frames similar to the key frames can be enhanced at receiver’s side, and thus they can be sent with lower quality. The transmitting device may determine frames being similar for example on the basis of one or more of the following: frames residing in the same shot, frames using the same coding mode (e.g. intra or inter), frames using the same quantization parameter, frames using the same lambda value or alike for rate-distortion optimized mode selections. The transmitting device uses the key frames for measuring a similarity measure with the currently captured frames. If they are similar enough in terms of a similarity threshold, the frames are streamed at lower quality, otherwise they will be streamed at higher quality. It is noted that performing a comparison is much less demanding in terms of required computation capability, memory or processing power than training a neural network.

[0221] This embodiment may be further modified, for example as follows:

The key frame sent from the receiving device to the transmitting device is an encoded version of a frame (hereafter referred to as frame N) that has been enhanced by the neural network. Alternatively to sending a key frame, a feature vector is sent from the receiving device to the transmitting device, wherein the feature vector is derived from frame N that has been enhanced by the neural network.

The transmitting device keeps the original uncompressed frame N in its storage. Such storing may be facilitated e.g. by selecting frame N in a periodic manner, where the period may be pre-defined or negotiated to be the same in the transmitting and receiving devices.

If the transmitting device receives an encoded key frame N from the receiving device, the transmitting device decodes the encoded key frame N and compares it to the original uncompressed frame N, e.g. using a distortion metric, such as block-wise sum of absolute differences. If the transmitting device receives a feature vector of frame N from the receiving device, the transmitting device derives a feature vector from the original uncompressed frame N (or a high-quality encoded and reconstructed version of frame N) and compares the feature vectors.

Based on either of the above comparisons, the transmitting device concludes whether the neural network-based enhancement of frame N was satisfactory (i.e. not causing undesired distortion). If the neural network-based enhancement was satisfactory, the transmitting device may conclude that it continues to encode frames of the same shot at low-quality streaming mode. Otherwise, the transmitting device may encode frames of the same shot at high quality and notify the receiving device to reset the neural network and/or to start fine-tuning the neural net again. Thus, it can be ensured that the enhancement does not accidentally cause significant distortions that were not present in the original captured video.

[0222] According to an embodiment, the transmitting device performs a shot boundary detection for the video data stream to be transmitted. Herein, the shot boundary detection performed by transmitting device may be a lightweight shot boundary detection. When the transmitting device detects a new shot, it may transmit the first few frames with high quality “just in case”, meaning that this is a safe choice to guarantee that the receiving device is always able to obtain high quality frames, i.e., there will be no quality drop for the first few frames of a previously unseen shot.

[0223] Further aspects of the invention relate to the operation of the second apparatus (the transmitting device). In accordance with what has been disclosed above, the operation of the second apparatus may be defined by a method comprising: transmitting, by a second apparatus, a first portion of a video data stream at a high quality to a first apparatus; receiving, from the first apparatus, one or more key frames of the first portion of a video data stream; and transmitting a second portion of said video data stream comprising frames sufficiently similar to the key frames at a lower quality to the first apparatus.

[0224] In general, the various embodiments of the invention may be implemented in hardware or special purpose circuits, software, logic or any combination thereof. For example, some aspects may be implemented in hardware, while other aspects may be implemented in firmware or software which may be executed by a controller, microprocessor or other computing device, although the invention is not limited thereto. While various aspects of the invention may be illustrated and described as block diagrams, flow charts, or using some other pictorial representation, it is well understood that these blocks, apparatus, systems, techniques or methods described herein may be implemented in, as non-limiting examples, hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controller or other computing devices, or some combination thereof.

[0225] The embodiments of this invention may be implemented by computer software executable by a data processor of the mobile device, such as in the processor entity, or by hardware, or by a combination of software and hardware. Further in this regard it should be noted that any blocks of the logic flow as in the Figures may represent program steps, or interconnected logic circuits, blocks and functions, or a combination of program steps and logic circuits, blocks and functions. The software may be stored on such physical media as memory chips, or memory blocks implemented within the processor, magnetic media such as hard disk or floppy disks, and optical media such as for example DVD and the data variants thereof, CD.

[0226] The memory may be of any type suitable to the local technical environment and may be implemented using any suitable data storage technology, such as semiconductor-based memory devices, magnetic memory devices and systems, optical memory devices and systems, fixed memory and removable memory. The data processors may be of any type suitable to the local technical environment, and may include one or more of general purpose computers, special purpose computers, microprocessors, digital signal processors (DSPs) and processors based on multi-core processor architecture, as non-limiting examples.

[0227] Embodiments of the inventions may be practiced in various components such as integrated circuit modules. The design of integrated circuits is by and large a highly automated process. Complex and powerful software tools are available for converting a logic level design into a semiconductor circuit design ready to be etched and formed on a semiconductor substrate.

[0228] Programs, such as those provided by Synopsys, Inc. of Mountain View, California and Cadence Design, of San Jose, California automatically route conductors and locate components on a semiconductor chip using well established rules of design as well as libraries of pre-stored design modules. Once the design for a semiconductor circuit has been completed, the resultant design, in a standardized electronic format (e.g., Opus, GDSII, or the like) may be transmitted to a semiconductor fabrication facility or "fab" for fabrication.

[0229] The foregoing description has provided by way of exemplary and non-limiting examples a full and informative description of the exemplary embodiment of this invention. However, various modifications and adaptations may become apparent to those skilled in the relevant arts in view of the foregoing description, when read in conjunction with the accompanying drawings and the appended claims. However, all such and similar

modifications of the teachings of this invention will still fall within the scope of this invention.

Claims

CLAIMS:

1. A method comprising

receiving, in a first apparatus, a first portion of a video data stream at a first quality;

using the high quality first portion of the video data stream to train a neural network for enhancing a video data stream of a type sufficiently similar to the first portion of the video data stream;

receiving, in the first apparatus, a second portion of said video data stream at a second quality, wherein the second quality is lower than the first quality; and

enhancing the quality of the second portion of said video data stream using the trained neural network.

2. The method according to claim 1, further comprising

processing the received first portion of the video data stream into a lower quality; inputting the processed first portion of the video data stream in said neural network;

computing a loss of the neural network on the basis of a difference between a video data stream at an output of the neural network and the received first portion of the video data stream;

deriving a training signal for the neural network based on the loss; and training the neural network with the training signal to decrease the loss.

3. The method according to claim 2, further comprising

training the neural network iteratively with an iteratively updated training signal until the loss meets at least one predetermined stopping criterion.

4. An apparatus comprising

means for receiving a first portion of a video data stream at a first quality;

means for using the first portion of the video data stream to train a neural network for enhancing a video data stream of a type sufficiently similar to the first portion of the video data stream;

means for receiving a second portion of said video data stream at a second quality, wherein the second quality is lower than the first quality; and means for enhancing the quality of the second portion of said video data stream using the trained neural network.

5. The apparatus according to claim 4, further comprising

means for processing the received first portion of the video data stream into a lower quality;

means for inputting the processed lower quality first portion of the video data stream in said neural network;

means for computing a loss of the neural network on the basis of a difference between a video data stream at an output of the neural network and the received first portion of the video data stream;

means for deriving a training signal for the neural network based on the loss; and means for training the neural network with the training signal to decrease the loss.

6. The apparatus according to claim 5, further comprising

means for training the neural network iteratively with an iteratively updated training signal until the loss meets at least one predetermined stopping criterion.

7. The apparatus according to any of claims 4 - 6, further comprising

means for analyzing, in the first apparatus, the received video data stream; and means for dividing the received video data stream into the first and any subsequent portions.

8. The apparatus according to claim 7, further comprising

means for identifying scene boundaries in the received video data stream; and means for dividing the received video data stream temporally into the first and any subsequent portions at locations of identified scene boundaries.

9. The apparatus according to claim 7 or 8, further comprising:

means for transmitting, responsive to determining that the predetermined stopping criterion is met, to a second apparatus a request to provide the second portion at the second quality.

10. The apparatus of according to any of claims 4 - 9, wherein the means comprises at least one processor and at least one memory, said at least one memory stored with code thereon, which when executed by said at least one processor, causes the performance of the apparatus.

11. A method comprising:

transmitting, by a second apparatus, a first portion of a video data stream at a first quality to a first apparatus;

receiving, from the first apparatus, one or more key frames, or parts thereof, or feature vectors thereof, wherein the one or more key frames correspond to a subset of the first portion of a video data stream; and

transmitting a second portion of said video data stream comprising frames sufficiently similar to the key frames at a second quality to the first apparatus, wherein the second quality is lower than the first quality.

12. An apparatus comprising:

means for transmitting a first portion of a video data stream at a first quality to a remote apparatus;

means for receiving, from the remote apparatus, one or more key frames, or parts thereof, or feature vectors thereof, wherein the one or more key frames correspond to a subset of the first portion of a video data stream; and

means for transmitting a second portion of said video data stream comprising frames sufficiently similar to the key frames at a second quality to the remote apparatus, wherein the second quality is lower than the first quality.

13. The apparatus according to claim 12, wherein sufficient similarity of the frames of the second portion to the key frames is configured to be determined on the basis of one or more of the following:

the frames of the second portion and the key frames residing in the same shot;

the frames of the second portion and the key frames using the same coding mode;

the frames of the second portion and the key frames using the same quantization parameter; the frames of the second portion and the key frames using the same lambda value for rate-distortion optimized mode selections.

14. The apparatus according to claim 12 or 13, wherein the one or more key frames are encoded versions of frames that have been enhanced by a neural network in the first apparatus.

15. The apparatus of according to any of claims 12 - 14, wherein the means comprises at least one processor and at least one memory, said at least one memory stored with code thereon, which when executed by said at least one processor, causes the performance of the apparatus.