RU2481727C2 - Advanced interpolation techniques for motion compensation in video coding - Google Patents

Abstract

FIELD: radio engineering, communication.

SUBSTANCE: encoder interpolates pixel values of reference video data based on a plurality of different predefined interpolation filters. The decoder receives a syntax element which identifies an interpolation filter, and interpolates pixel values of reference video data based on the interpolation filter identified by the syntax element. In another example, a method of interpolating predictive video data includes generating half-pixel values based on integer pixel values, rounding the half-pixel values to generate half-pixel interpolated values, storing the half-pixel values as non-rounded versions of the half-pixel values and generating quarter-pixel values based on the non-rounded versions of the half-pixel values and the integer pixel values.

EFFECT: interpolation technique which is executable by an encoder and a decoder during the motion compensation process of video coding.

26 cl, 11 dwg, 8 tbl

Description

This application claims the priority of provisional patent application (US) 61/044023, filed April 10, 2008, provisional patent application (US) 61/044240, filed April 11, 2008, and provisional patent application (US) number 61 / 057373, filed May 30, 2008, the entire contents of which are contained in this document by reference.

FIELD OF THE INVENTION

This disclosure relates to digital video coding, and more specifically to fractional prediction data interpolations used in video coding.

BACKGROUND

Digital video capabilities can be embedded in a wide range of devices, including digital TVs, digital direct broadcast systems, wireless broadcasting systems, personal digital devices (PDAs), portable or desktop computers, digital cameras, digital recording devices, video game devices, video game consoles, cellular or satellite radiotelephones, etc. Digital video devices implement video compression technologies such as MPEG-2, MPEG-4, or ITU-T H.264 / MPEG-4, part 10, advanced video coding (AVC) to more efficiently transmit and receive digital video information. Video compression technologies may perform spatial prediction and / or temporal prediction to reduce or remove redundancy internally present in video sequences.

Block-based cross-coding is a very useful coding technology that is based on temporal prediction in order to reduce or remove temporal redundancy between video blocks of consecutive coded units of a video sequence. Coded units may comprise video frames, a series of consecutive macroblocks of video frames, image groups, or another specified unit of video blocks. For inter-coding, the video encoder performs motion estimation in order to track the movement of respective video blocks between two or more adjacent coded units. The motion estimation generates motion vectors that indicate the displacement of the video blocks relative to the corresponding predicted video blocks in one or more reference frames or other coded units. Motion compensation uses motion vectors to form predictive video blocks from a reference frame or other coded unit. After motion compensation, residual video blocks are formed by subtracting the predicted video blocks from the encoded source video blocks.

The video encoder may also apply transform, quantization, and entropy coding processes to further reduce the bit rate associated with the transmission of residual blocks. Transformation technologies may include discrete cosine transforms (DCT) or conceptually similar processes. Alternatively, wavelet transforms, integer transforms, or other types of transforms may be used. In the DCT process, as an example, a set of pixel values is converted into transform coefficients, which can represent the energy of the pixel values in the frequency domain. Quantization is applied to transform coefficients and, in general, encompasses a process that limits the number of bits associated with any given transform coefficient. Entropy coding comprises one or more processes that together compress a sequence of quantized transform coefficients. Examples of entropy coding include, but are not limited to, context-adaptive variable-length coding (CAVLC) and context-adaptive binary arithmetic coding (CABAC).

An encoded video block may be represented by prediction information that can be used to create or identify a prediction block and a residual data block indicating differences between the encoded block and the prediction block. The prediction information may comprise one or more motion vectors that are used to identify the predictive data block. Given the motion vectors, the decoder can reconstruct the prediction blocks that are used to encode the remainder. Thus, given the set of residual blocks and the set of motion vectors (and possibly some additional syntax), the decoder may be able to recover the video frame that was originally encoded. Mutual coding based on motion estimation and motion compensation makes it possible to achieve very good compression, since sequential video frames or other types of coded units are often pretty much the same. An encoded video sequence may contain blocks of residual data, motion vectors, and possibly other types of syntax.

Interpolation technologies are designed to increase the level of compression that can be achieved with cross-coding. In this case, the predictive data generated during the motion compensation, which is used to encode the video block, can be interpolated from the pixels of the video blocks of the video frame or other coded unit used in the motion estimation. Interpolation is often performed in order to generate predictive half-pixel values (at half a pixel) and predictive quarter-pixel values (at a quarter of a pixel). Such interpolation often forms predictive blocks, which are even more similar to the encoded video blocks than the actual video blocks of the predictive frame or other coded unit used in video encoding.

SUMMARY OF THE INVENTION

In general, this disclosure describes interpolation techniques performed by an encoder and a decoder during a video encoding motion compensation process. According to the techniques of this disclosure, an encoder may apply a plurality of predetermined interpolation filters to generate a plurality of different interpolated prediction data. Interpolated predictive data that reaches the highest levels of compression can be selected, and the interpolation filter that is used can be encoded as syntax and transmitted to the decoding device as part of the encoded video data. An interpolation filter can be set once per predictive unit (for example, once per predictive frame) or can be set based on blocks (for example, different filtering can be applied to different video blocks within a predictive unit). Alternatively, the interpolation filter may be specified based on samples or may be defined based on subsamples.

The decoder receives encoded video data and can interpret the syntax to identify the interpolation filter that is used by the encoder. Thus, the decoder can identify and use identical interpolation filters during the motion compensation processes that are used during the encoding process. By considering several different interpolation filters in an encoder, compression can be improved relative to technologies that have a non-tunable interpolation filter. However, technologies can be significantly less complex than traditional adaptive interpolation filtering, which adaptively sets interpolation filtering coefficients as part of the encoding process. Various interpolation filters, according to some aspects of this disclosure, can be predefined by implementing adaptive interpolation filtering or other technologies for test video sequences to pre-define sets of interpolation filter coefficients that are likely to lead to good compression. Alternatively, different numbers of filter coefficients (different numbers of taps) or different types of filters, in general, can be predefined and then selected and used during the encoding and decoding processes.

In addition, this disclosure also recognizes coding inefficiencies due to traditional half-pixel rounding and provides technologies that can improve interpolation by reducing or eliminating intermediate rounding. In this case, the interpolated half-pixel values may be rounded for the purpose of half-pixel interpolation. However, quarter-pixel values that can be generated based on one or more interpolated half-pixel values can be based on unrounded versions of half-pixel values. This eliminates the spread of rounding inaccuracies from half-pixel values to quarter-pixel values. In some cases, slight rounding without sacrificing the accuracy of the final values can be applied to one particular half-pixel value to ensure that sixteen-bit storage elements can be used to store all intermediate half-pixel values. In particular, when fifteen possible sub-pixel locations are specified for each pixel location, one of the specific half-pixel values may need to be generated based on the other half-pixel values (i.e., requiring two levels of half-pixel interpolation), and this particular half-pixel value may require rounding in order to ensure that sixteen-bit storage elements can be used to store all interpolated values.

In any case, the elimination of the tasks of intermediate rounding can be carried out with or without the help of implementing a plurality of predetermined interpolation filters, as described in this document. For example, the elimination of intermediate rounding may be performed in the context of an implementation that uses a plurality of predefined interpolation filters, as described herein. Alternatively, the elimination of intermediate rounding can be performed in the context of more traditional interpolations, for example, in the context of adaptive interpolation filtering (AIF).

In one example, this disclosure provides a method of encoding video data. The method comprises generating predictive data, and generating the predictive data includes interpolating pixel values of the reference video data based on a plurality of different predetermined interpolation filters. The method also comprises encoding video data based on predictive data.

In another example, this disclosure provides a method for decoding video data. The decoding method comprises receiving a syntax element that identifies an interpolation filter from a plurality of different predetermined interpolation filters, generating predictive data, and generating the predictive data includes interpolating pixel values of the reference video data based on an interpolation filter identified by the syntax element and decoding the video data to Based on predictive data.

In another example, this disclosure provides a method for interpolating predictive video data for encoding a video. The method comprises generating half-pixel values based on integer-pixel values, rounding half-pixel values to generate half-pixel interpolated values, storing half-pixel values as unrounded versions of half-pixel values, and generating quarter-pixel values based on unrounded versions of half-pixel values and integer-pixel values.

In another example, this disclosure provides a device that encodes video data, the device comprising a video encoder that includes a motion compensation module that generates predictive data. The motion compensation module interpolates the pixel values of the reference video data based on a plurality of different predetermined interpolation filters, and the video encoder encodes the video data based on the predicted data.

In another example, this disclosure provides a device that decodes video data, the device comprising a video decoder that includes a motion compensation module. The video decoder receives a syntax element that identifies an interpolation filter from a plurality of different predetermined interpolation filters. The motion compensation module generates predictive data, while generating the predictive data includes interpolating the pixel values of the reference video data based on an interpolation filter identified by a syntax element. The video decoder then decodes the video data based on the predictive data.

In another example, this disclosure provides a device that interpolates predictive video data for encoding video, the device includes a motion compensation module that generates half-pixel values based on integer-pixel values, rounds half-pixel values to generate half-pixel interpolated values, stores half-pixel values as unrounded versions of half-pixel values and generates quarter-pixel values based on unrounded versions of olupixel values and integer-pixel values.

In another example, this disclosure provides a device that encodes video data, the device comprising means for generating predictive data, the means for generating predictive data including means for interpolating pixel values of the reference video data based on a plurality of different predetermined interpolation filters and means for encoding video data based on predictive data.

In another example, this disclosure provides a device that decodes video data, the device comprising means for receiving a syntax element that identifies an interpolation filter from a plurality of different predetermined interpolation filters, means for generating predictive data, the means for generating predictive data including means for interpolating pixel values of the reference video data based on an interpolation filter identified in the middle Twomey syntax element, and means for decoding the video data based on the predictive data.

In another example, this disclosure provides a device that interpolates predictive video data for encoding video, the device comprising means for generating half-pixel values based on integer-pixel values, means for rounding half-pixel values to form half-pixel interpolated values, means for storing half-pixel values as non-rounded versions of half-pixel values and a means for generating quarter-pixel values based on neocr Glen versions of the half pixel values and tselopikselnyh values.

In another example, this disclosure provides a device that encodes video data, the device comprising a video encoder including a motion compensation module that generates predictive data by interpolating pixel values of the reference video data based on a plurality of different predetermined interpolation filters and encodes video data based on the predictive data, and a wireless transmitter that transmits video data to another device.

In another example, this disclosure provides a device that decodes video data, the device comprising a wireless receiver that receives video data, and a video decoder including a motion compensation module that receives a syntax element that identifies an interpolation filter from a plurality of different predetermined interpolation filters, generates predictive data, while the formation of predictive data includes the interpolation of pixel values PORN video data based on the interpolation filter identified by the syntax element, and decodes the video data based on predictive data.

In another example, this disclosure provides a device that encodes video data, the device comprising a video encoder including a motion compensation module that generates half-pixel values based on integer-pixel values, rounds half-pixel values to generate half-pixel interpolated values, and stores half-pixel values as non-rounded versions of half-pixel values and generates quarter-pixel values based on unrounded versions of half-pixel values values and integer-pixel values, and a wireless transmitter that transmits video encoded by the motion compensation module to another device.

In another example, this disclosure provides a device that decodes video data, the device comprising a wireless receiver that receives video data, and a video decoder including a motion compensation module that decodes video data. When decoding video data, the motion compensation module generates half-pixel values based on integer-pixel values, rounds half-pixel values to generate half-pixel interpolated values, saves half-pixel values as unrounded versions of half-pixel values, and generates quarter-pixel values based on unrounded versions of half-pixel values and integer-pixel values.

The technologies described in this disclosure may be implemented in hardware, software, firmware, or any combination of the above. If implemented in software, the software may run on one or more processors, such as a microprocessor, application specific integrated circuit (ASIC), field programmable gate array (FPGA), or digital signal processor (DSP). Software that implements technology can be initially stored on a machine-readable medium and downloaded and executed on a processor.

Accordingly, this disclosure also considers a computer-readable storage medium containing instructions that, when executed, instruct the device to encode video data, while instructions instruct the device to generate predictive data, while generating predictive data includes interpolating pixel values of the reference video data based on a plurality of different advance preset interpolation filters and video coding based on predictive data.

This disclosure also considers a computer-readable storage medium containing instructions that, when executed, instruct the device to decode video data, while the instructions instruct the device, upon receiving a syntax element that identifies an interpolation filter from a plurality of different predetermined interpolation filters, to generate predictive data, while generating predictive data includes interpolating the pixel values of the reference video data based on nterpolyatsionnogo filter identified by the syntax element, and decode the video data based on predictive data.

In addition, this disclosure also considers a computer-readable storage medium containing instructions that, when executed, instruct the device to interpolate predictive video data for video encoding, while instructions instruct the device to generate half-pixel values based on integer-pixel values, round off half-pixel values to form half-pixel interpolated values, save half-pixel values as unrounded versions of half-pixel values and generate quarter pixel values based on unrounded versions of half pixel values and integer pixel values.

Details of one or more embodiments of this disclosure are set forth in the accompanying drawings and in the description below. Other features, objectives and advantages of the technologies described in this disclosure should be apparent from the description and drawings, as well as from the claims.

BRIEF DESCRIPTION OF THE DRAWINGS

1 is an exemplary block diagram illustrating a video encoding and decoding system.

2 is a block diagram illustrating an example of a video encoder in accordance with this disclosure.

3 is a block diagram illustrating an example of a video decoder in accordance with this disclosure.

4A-4D are conceptual diagrams illustrating integer pixels and interpolated pixels that may contain pixels of predictive video blocks.

5-8 are flowcharts illustrating techniques in accordance with this disclosure.

DETAILED DESCRIPTION OF THE INVENTION

This disclosure describes various interpolation techniques performed by an encoder and a decoder during a video encoding motion compensation process. According to one aspect of this disclosure, an encoder may apply a plurality of predetermined interpolation filters to generate a plurality of different interpolated prediction data. Interpolated predictive data that reaches the highest levels of compression can be selected in the encoder, and the interpolation filter that is used can be encoded as syntax and transmitted to the decoding device as part of the encoded video data. The interpolated predictive data contains reference data. The video data to be encoded (e.g., a video block) can be subtracted from the interpolated predictive data (e.g., an interpolated predictive reference block) to define a residual block of encoded data, which can then be transformed, quantized, and entropy encoded.

An interpolation filter can be set once per predictive unit (for example, once per predictive frame) or can be set based on samples (for example, different filtering can be applied to different video blocks or different approximate locations within the predictive unit ) Different interpolation filters, for example, can be defined by different sets of filter coefficients, different numbers of filter coefficients, or possibly different types of filters. In any case, by providing a plurality of predetermined interpolation filters, the interpolated data may provide for improved data compression.

The decoder receives encoded video data and can interpret the syntax to identify the interpolation filter that is used by the encoder. In this way, the decoder can identify and use the interpolation filter during its motion compensation process, which is identical to the filter used during the encoding process. On the other hand, by considering several different interpolation filters in the encoder, compression can be improved with respect to technologies that have one predetermined interpolation filter. However, technologies can be significantly less complex than traditional adaptive interpolation filtering, which adaptively defines an interpolation filter (by selecting filter coefficients) as part of the encoding process. In one example, interpolation filters can be predefined by implementing adaptive interpolation filtering or other technologies for test video sequences to pre-define sets of interpolation filter coefficients that are likely to lead to good compression.

In another aspect of this disclosure, interpolation techniques are described that eliminate or significantly reduce the intermediate rounding of half-pixel values for quarter-pixel interpolation. Technologies that eliminate or significantly reduce intermediate rounding can be used with or without the technologies mentioned above that use predefined interpolation filters. In other words, technologies that eliminate or significantly reduce intermediate rounding can be used during interpolation according to one or more predefined interpolation filters, but can also be used for interpolation that uses a non-tunable interpolation filter, or possibly adaptive refinement interpolation technologies.

This disclosure recognizes coding inefficiencies that may occur due to traditional rounding of half-pixel values, and proposes technologies to improve interpolation by eliminating or reducing intermediate rounding. In this case, the interpolated half-pixel values may be rounded for the purpose of half-pixel interpolation. However, quarter-pixel values that can be generated based on one or more interpolated half-pixel values can be based on unrounded versions of half-pixel values. This eliminates the spread of rounding inaccuracies from half-pixel values to quarter-pixel values. In some cases, slight rounding may be applied to some of the half-pixel values to ensure that a sixteen-bit data structure can be used to store all intermediate values. In some cases, the effect of small rounding on the final accuracy may be zero by properly rounding. In any case, the elimination or reduction of the intermediate rounding can be carried out with or without the help of implementing a plurality of predetermined interpolation filters, as described herein.

1 is a block diagram illustrating one exemplary video encoding and decoding system 10 that may implement the techniques of this disclosure. As shown in FIG. 1, system 10 includes a source device 12 that transmits encoded video to a target device 16 via a communication channel 15. The source device 12 and the target device 16 may include any of a wide range of devices. In some cases, the source device 12 and the target device 16 comprise wireless communication devices, such as wireless portable telephones, so-called cellular or satellite radiotelephones, or any wireless devices that can transmit video information over communication channel 15 when communication channel 15 is wireless. The technologies of this disclosure, however, which relate to motion-compensated interpolation, are not necessarily limited to wireless applications or settings.

In the example of FIG. 1, the source device 12 may include a video source 20, a video encoder 22, a modulator / demodulator (modem) 23, and a transmitting device 24. The target device 16 may include a receiving device 26, a modem 27, a video decoder 28, and a display device 30. According to this disclosure, video encoder 22 of source device 12 may be configured to apply one or more interpolation technologies as part of a video encoding process. Similarly, the video decoder 28 of the target device 16 may be configured to apply one or more interpolation technologies of this disclosure as part of the video decoding process.

The illustrated system 10 of FIG. 1 is merely exemplary. The interpolation technologies of this disclosure may be performed by any encoding device that supports motion-compensated interpolation up to sub-pixel resolution. The source device 12 and the target device 16 are simply examples of such encoding devices. In this case, the source device 12 generates encoded video data for transmission to the target device 16. The devices 12, 16 can operate almost symmetrically so that each of the devices 12, 16 includes video encoding and decoding components, for example, in a combined encoder-decoder (codec). Therefore, the system 10 can support one-way and two-way video transmission between video devices 12, 16, for example, for video streaming, video playback, video broadcasting, or video telephony.

The video source 20 of the source device 12 may include a video capture device, such as a video camera, a video archive containing previously captured video, or a video stream from a video content provider. As an additional alternative, video source 20 may generate computer-based data as a source video or a combination of direct-transmitted video, archived video, and machine-generated video. In some cases, if the video source 20 is a video camera, the source device 12 and the target device 16 can form so-called camera phones or video phones. In each case, the captured, pre-captured, or machine-generated video may be encoded by video encoder 22. The encoded video information may then be modulated by modem 23 according to a communication standard, such as, for example, code division multiple access (CDMA) or another communication standard, and transmitted to the target device 16 via the transmitting device 24. The modem 23 may include various mixers, filters, amplifiers, or other components designed to modulate the signal at. For wireless applications, transmitter 24 may include circuits designed for transmitting data, including amplifiers, filters, and one or more antennas.

The receiver 26 of the target device 16 receives information on channel 15, and the modem 27 demodulates the information. Alternatively, the video encoding process may implement one or more of the techniques described herein to improve interpolation during motion compensation. The video decoding process performed by the video decoder 28 can also perform interpolation during the motion compensation stage of the decoding process. The display device 30 displays the decoded video data to the user and may include any of a variety of display devices, such as a cathode ray tube, liquid crystal display (LCD), plasma display, organic light emitting diode (OLED) display, or other type of display device.

In the example of FIG. 1, communication channel 15 may comprise any wireless and wired communication medium, such as a radio frequency (RF) spectrum or one or more physical transmission lines, or any combination of wireless and wired media. Communication channel 15 may form part of a packet-switched network, such as a local area network, wide area network, or wide area network such as the Internet. Communication channel 15 generally represents any suitable communication medium or a set of different communication media for transmitting video data from source device 12 to target device 16.

Video encoder 22 and video decoder 28 may operate according to a video compression standard, such as ITU-T H.264, alternatively described as MPEG-4, part 10, advanced video encoding (AVC). The technologies of this disclosure are, however, not limited to any particular coding standard. Although not shown in FIG. 1, in some aspects, video encoder 22 and video decoder 28 may be integrated with an audio encoder and decoder, respectively, and may include appropriate multiplexer-demultiplexer modules or other hardware and software to process the encoding as audio video and in the general data stream or in separate data streams. If applicable, the multiplexer-demultiplexer modules may conform to the ITU H.223 multiplexer protocol or other protocols, such as user datagram protocol (UDP).

The ITU H.264 / MPEG-4 (AVC) standard is formulated by the ITU-T Video Coding Expertise Group (VCEG) ITU-T in conjunction with the ISO / IEC Motion Picture Expert Group (MPEG) as a joint partnership project known as the Joint Video Standards Group ( JVT). In some aspects, the techniques described in this disclosure may be applied to devices that, in general, comply with the H.264 standard. The H.264 standard is described in ITU-T Recommendation H.264 "Advanced Video Coding for Generic Audiovisual Services" from the ITU-T research group and is dated March 2005, which may be referred to in this document as the H.264 standard or the H.264 specification either a standard or an H.264 / AVC specification. The Joint Video Standards Group (JVT) continues to work on additions to the H.264 / MPEG-4 AVC.

Video encoder 22 and video decoder 28 can be implemented as one or more microprocessors, digital signal processors (DSPs), specialized integrated circuits (ASICs), field programmable gate arrays (FPGAs), discrete logic, software, hardware, firmware, or any combination of of the foregoing. Each of video encoder 22 and video decoder 28 can be included in one or more encoders or decoders, any of which can be integrated as part of a combined codec in a corresponding mobile device, subscriber device, broadcast device, server, and the like.

A video sequence typically includes a series of video frames. Video encoder 22 operates with video blocks within separate video frames to encode video data. Video blocks may have a fixed or variable size and may vary in size according to a given coding standard. Each video frame includes a series of consecutive macroblocks. Each series of consecutive macroblocks may include a series of macroblocks that can be arranged in subunits. As an example, the ITU-T H.264 standard supports intra prediction for various block sizes, such as 16x16, 8x8 or 4x4 for luminance signal components and 8x8 for chroma signal components, as well as inter prediction for various block sizes, such as 16x16. 16x8, 8x16, 8x8, 8x4, 4x8 and 4x4 for luminance components and corresponding scaled sizes for chroma components. Video blocks may comprise blocks of pixel data or blocks of transform coefficients, for example, after a transform process, such as a discrete cosine transform or a conceptually similar transform process.

Smaller video blocks can provide better resolution and can be used for video frame locations that include high levels of detail. In general, macroblocks (MBs) and various subblocks can be considered video blocks. In addition, a series of consecutive macroblocks can be considered a series of video blocks, such as MB and / or subunits. Each series of consecutive macroblocks can be an independently decoded video frame unit. Video encoder 22 and video decoder 28 perform mutual predictive coding, which involves generating predictive reference data and subtracting the video block to be encoded from predictive reference data to generate residual data, which can then be converted, quantized, and entropy encoded. Mutual predictive coding may include interpolation of predictive data in accordance with this disclosure.

After mutual predictive coding (which includes the interpolation technology of this disclosure) and after all transformations (for example, the integer 4x4 or 8x8 transform used in H.264 / AVC, or the discrete cosine transform, DCT), quantization can be performed. Quantization generally means a process in which coefficients are quantized to possibly reduce the amount of data used to represent the coefficients. The quantization process may reduce the bit depth associated with some or all of the coefficients. For example, a 16-bit value can be rounded down to a 15-bit value during quantization. After quantization, entropy coding can be performed, for example, according to context-adaptive variable-length coding (CAVLC), context-adaptive binary arithmetic coding (CABAC), or other entropy coding technology.

According to the techniques of this disclosure, video encoder 22 may apply a plurality of predetermined interpolation filters to generate a plurality of different interpolated prediction data. Interpolated predictive data that reaches the highest compression levels can be selected by video encoder 22, and the interpolation filter that is used for the selected interpolated predictive data can be encoded as syntax and transmitted to the target device 16 as part of the encoded video data. In this case, the video decoder 28 receives the encoded video data and can interpret the syntax to identify the interpolation filter that is used by the video encoder 22. Accordingly, the video decoder 28 can identify and use the interpolation filter during its motion-compensated decoding process, which is identical to the filter used in during the coding process. Different predefined interpolation filters may contain similar filters with different sets of filter coefficients (filter taps). Alternatively, various predetermined interpolation filters may comprise filters having different numbers of filter taps or, possibly, different types of filter configurations in general.

In addition, according to another aspect of this disclosure, video encoder 22 and video decoder 28 may employ interpolation techniques that eliminate or significantly reduce the intermediate rounding of half-pixel values for quarter-pixel interpolation. In this case, video encoder 22 and video decoder 28 may round off interpolated half-pixel values for half-pixel interpolation. For quarter-pixel interpolation, however, video encoder 22 and video decoder 28 can save and use non-rounded versions of half-pixel values to reduce or eliminate the spread of rounding inaccuracies from half-pixel values to quarter-pixel values. In some cases, slight rounding can be applied to one particular half-pixel value, which requires two levels of interpolation to ensure that fixed-size storage elements (e.g., sixteen-bit registers) can be used to store all intermediate values. In some cases, the effect of small rounding on the final accuracy can be set to zero by properly rounding as described in this document.

FIG. 2 is a block diagram illustrating an example of a video encoder 50 that can perform motion compensation interpolation techniques in accordance with this disclosure. Video encoder 50 may correspond to video encoder 22 of device 20 or a video encoder of another device. Video encoder 50 may perform inter-coding and inter-coding of blocks within video frames, although inter-coding is not illustrated. Intra-coding is based on spatial prediction to reduce or remove spatial redundancy in a given video frame. Mutual coding is based on temporal prediction to reduce or remove temporal video redundancy in adjacent frames of a video sequence. The internal mode (I-mode) may be referred to as spatial compression mode, and mutual modes, such as the prediction mode (P-mode) or bidirectional mode (mode B), may be referred to as temporary compression modes. The techniques of this disclosure may typically be applied during inter-coding, and therefore, internal coding modules, for example, a spatial prediction module, are not illustrated in FIG. 2 for simplicity and convenience of illustration. However, the rounding technologies of this disclosure may also be applicable to spatial prediction and internal coding technologies.

As shown in FIG. 2, video encoder 50 receives the current video block in a video frame to be encoded. In the example of FIG. 2, video encoder 50 includes a motion estimation module 32, a motion compensation module 35, a reference frame storage module 34, an adder 48, a transform module 38, a quantization module 40, and an entropy encoding module 46. To restore the video block, video encoder 50 also includes an inverse quantization module 42, an inverse transform module 44, and an adder 51. A deblocking filter (not shown) may also be included to filter block boundaries to remove blocking artifacts from the reconstructed video. If desired, the deblocking filter typically should filter the output of the adder 51.

During the encoding process, video encoder 50 receives a video block to be encoded, and motion estimation module 32 and motion compensation module 35 perform mutual predictive encoding. Motion estimation module 32 and motion compensation module 35 may have a high degree of integration, but are illustrated separately for conceptual purposes. Motion estimation is typically considered the process of generating motion vectors that evaluate motion for video blocks. The motion vector, for example, may indicate the offset of the prediction block within the prediction frame (or other coded unit) relative to the current block encoded within the current frame (or other coded unit). Motion compensation is typically considered a process of sampling or generating a prediction block based on a motion vector determined by motion estimation. On the other hand, motion estimation module 32 and motion compensation module 35 can be functionally integrated. The interpolation techniques described in this disclosure are described as being performed by the motion compensation module 35. However, interpolation can be performed during motion estimation to simplify the selection of the best motion vectors.

According to this disclosure, motion estimation 32 selects the corresponding motion vector for the video block to be encoded by comparing the video block with the video blocks of the predictive coded unit (e.g., the previous frame). Here, motion compensation module 35 may perform interpolation to generate predictive data at sub-pixel resolution. In some cases, during motion estimation, interpolation may be based on a non-tunable interpolation filter. In other cases, various interpolation filters used during motion compensation (as indicated below) can also be used during the motion estimation process to select a motion vector.

After the motion estimation module 32 selects a motion vector for the video block to be encoded, the motion compensation module 35 generates a predictive video block associated with this motion vector. According to this disclosure, however, motion compensation module 35 may consider several versions of any predictive video block that has sub-pixel resolution. In this case, the motion compensation module 35 may apply a plurality of predetermined interpolation filters to generate a plurality of different interpolated prediction data for the video block to be encoded. Motion compensation module 35 then selects the interpolated prediction data (for example, the interpolated video block associated with one of the interpolation filters) that achieve the highest compression levels. An interpolation filter, which is used to generate interpolated data, can be encoded as interpolation syntax and transmitted to entropy encoding unit 46 for inclusion in the encoded bitstream. After the motion compensation module 35 selects and applies the best interpolation filter, the motion compensation module 35 generates predictive data using this interpolation filter, and the video encoder uses an adder 48 to subtract these predictive data from the encoded video block to generate residual data.

In addition, according to another aspect of this disclosure, video encoder 22 and video decoder 28 may employ interpolation techniques that eliminate or significantly reduce the intermediate rounding of half-pixel values for quarter-pixel interpolation. In this case, video encoder 22 and video decoder 28 may round off interpolated half-pixel values for half-pixel interpolation. For quarter-pixel interpolation, however, video encoder 22 and video decoder 28 can save and use non-rounded versions of half-pixel values to reduce or eliminate the spread of rounding inaccuracies from half-pixel values to quarter-pixel values. In some cases, a slight rounding can be applied to one particular half-pixel value, which requires two levels of interpolation to ensure that fixed-size storage elements (e.g. 16-bit registers) can be used to store all intermediate values.

As noted, after the motion compensation module 35 generates predictive data (for example, an interpolated predictive video block), video encoder 50 generates a residual video block by subtracting the predictive data from the encoded source video block. Adder 48 represents the component or components that perform this subtraction operation. The transform unit 38 applies a transform, such as a discrete cosine transform (DCT) or a conceptually similar transform, to the residual block, forming a video block containing residual block transform coefficients. Conversion module 38, for example, may perform other transformations defined by the H.264 standard, which are conceptually similar to DCTs. Wavelet transforms, integer transforms, subband transforms, or other types of transforms can also be used. In any case, the transform unit 38 applies the transform to the residual block, forming a block of residual transform coefficients. The conversion may convert the residual information from the pixel domain into the frequency domain.

Quantization module 40 quantizes the residual transform coefficients to further reduce the bit rate. The quantization process may reduce the bit depth associated with some or all of the coefficients. For example, a 16-bit value can be rounded down to a 15-bit value during quantization. After quantization, entropy encoding unit 46 entropy encodes the quantized transform coefficients. For example, entropy encoding unit 46 may perform variable length context adaptive encoding (CAVLC), context adaptive binary arithmetic encoding (CABAC), or other entropy encoding technology. After entropy encoding by means of entropy encoding unit 46, the encoded video can be transferred to another device or archived for subsequent transmission or retrieval. The encoded bitstream may include entropy encoded residual blocks, motion vectors for such blocks, and other syntax, including interpolation syntax, which identifies interpolation filters that are applied by motion compensation module 35.

The inverse quantization unit 42 and the inverse transform unit 44 apply inverse quantization and inverse transform, respectively, to restore the residual block in the pixel domain, for example, for subsequent use as a reference block. The adder 51 adds the reconstructed residual block to the motion compensation predicted block generated by the motion compensation unit 35 to form the reconstructed video block for storage in the reference frame storage unit 34. The reconstructed video block may be used by the motion estimation module 32 and the motion compensation module 35 as a reference block in order to mutually encode the block in a subsequent video frame.

FIG. 3 is a block diagram illustrating an example of a video decoder 60 that decodes a video sequence that is encoded in the manner described herein. Video decoder 60 includes a motion compensation module 55 that performs interpolation techniques of this disclosure for decoding. In particular, on the decoding side, motion compensation module 55 may receive a syntax element from entropy decoding module 52, which identifies an interpolation filter from a plurality of different predetermined interpolation filters. Motion compensation module 55 may generate predictive data, which includes interpolating the pixel values of the reference video data based on an interpolation filter identified by a syntax element. In particular, motion compensation module 55 may generate predictive data based on motion vectors received from entropy decoding module 52 and interpolations as specified by a syntax element (labeled as interpolation syntax in FIG. 3). Based on these interpolated predictive data, video data (e.g., reconstructed residual video block) can be decoded.

Entropy decoding unit 52 entropy decodes the received bitstream to generate quantized coefficients and syntax (for example, interpolation syntax and motion vectors sent to motion compensation unit 55). The inverse quantization unit 56 inversely quantizes, i.e. quantizes block block coefficients. The inverse quantization process may be a conventional process as specified by H.264 decoding. The inverse transform unit 58 applies the inverse transform, for example, the inverse DCT or a conceptually similar inverse transform process, to the transform coefficients to form residual blocks in the pixel domain. Motion compensation module 55 generates motion compensation blocks in a manner described herein, for example, including interpolation based on a set of interpolation filtering coefficients identified by a syntax element (i.e., interpolation syntax).

The adder 64 decodes the residual blocks by summing the residual blocks with the corresponding prediction blocks generated by the motion compensation unit 55 to form the decoded blocks. If desired, a deblocking filter can also be applied to filter decoded blocks to remove blocking artifacts. The decoded video blocks are then stored in a reference frame storage unit 62, which is a storage element that provides reference blocks for subsequent motion compensation and also generates decoded video to drive a display device (such as the device 30 of FIG. 1).

On the other hand, the technologies of this disclosure relate to motion-compensated interpolation in which the pixel values of the predicted video blocks are interpolated to sub-pixel resolution. The encoder uses the techniques of this disclosure to identify a desired interpolation filter from a plurality of predetermined interpolation filters. Different filters may differ by different sets of filter coefficients, different numbers of filter coefficients, or different types of filters. The decoder interprets the syntax elements sent from the encoder to identify the identical desired set of interpolation filtering coefficients used by the encoder.

4A-4D are conceptual diagrams illustrating integer pixels and interpolated pixels that may contain pixels of predictive video blocks. In the conceptual illustration of FIG. 4, various rectangles represent pixels. Uppercase letters (in rectangles with solid lines) represent integer-pixel locations, while lowercase letters (in rectangles with dashed lines) represent interpolated pixel locations. Letter marks may be used herein to describe pixel locations or pixel positions, or may be referred to as pixel values associated with different locations. The pixel locations "aa", "bb", "cc", "dd", "ee", "ff", "gg", "hh", "ii" and "jj" are half-pixel locations used in fractional interpolations of various fractional locations associated with the pixel location "C3".

Each pixel location can have associated fifteen different fractional locations, for example, in accordance with interpolations compatible with the H.264 / AVC ITU standard. In the example of FIGS. 4A-4D, these 15 different fractional locations associated with the pixel “C3” are illustrated. For simplicity and convenience of illustration, most of the other fractional locations are not shown (except those mentioned above, which are used to form one or more of 15 different fractional locations associated with the pixel "C3").

In the ITU H.264 / AVC standard, in order to receive luminance signals in half-pixel positions, a Wiener 6-tap filter with coefficients [1, -5, 20, 20, -5, 1] is typically used. Then, to obtain luminance signals at quarter-pixel locations, a bilinear filter is used. The bilinear filter can also be used in fractional pixel interpolation for color signal components that can have an accuracy of up to 1/8 pixel in H.264 / AVC.

Half pixel interpolations of the pixel locations “b” and “h” are shown in FIG. 4B in the horizontal and vertical directions, respectively. In this case, the pixel location “b” can be interpolated based on the integer pixels “C1”, “C2”, “C3”, “C4”, “C5” and “C6”. Similarly, the pixel location “h” can be interpolated based on the integer pixels “A3”, “B3”, “C3”, “D3”, “E3” and “F3”. Different interpolation filters (for example, different sets of filter taps) can be used to generate different interpolated pixel locations "b" and "h", as described herein. 4B, the interpolated pixel locations “b” and “h” are shown by shading, and the integer pixel locations “C1”, “C2”, “C3”, “C4”, “C5”, “C6”, “A3”, “B3”, “C3”, “D3”, “E3” and “F3” are indicated by the signs “pound”.

Fig. 4C illustrates one particular case that may require two levels of interpolation. In particular, the pixel location “j” is different from other half-pixel locations, since the pixel location “j” itself is interpolated based on other half-pixel values. For example, the pixel location “j” can be interpolated based on the half-pixel interpolated values “cc”, “dd”, “h”, “ee”, “ff” and “gg” in the horizontal direction. Alternatively, the pixel location “j” can be interpolated based on the half-pixel interpolated values “aa”, “bb”, “b”, “hh”, “ii” and “jj” in the vertical direction. 4C, the interpolated pixel location "j" is shown by shading, the half-pixel interpolated values "aa", "bb", "b", hh "," ii "and" jj "in the vertical direction are shown by the signs" pound " from right to left, and the half-pixel interpolated values of "cc", "dd", "h", "ee", "ff" and "gg" are shown using the signs "pound" from left to right.

FIG. 4D illustrates quarter pixel locations of “a,” “c,” “d,” “e,” “f,” “g,” “i,” “k,” “l,” “m,” “n,” and “o” by shading and illustrates integer- and half-pixel locations used for such quarter-pixel interpolation (for example, “C3”, “b”, “C4”, “h”, “j”, “ee”, “D3” , "hh" and "D4") using the pound signs.

After motion estimation, the best motion vector for a given video block can be identified, for example, possibly using a distortion model depending on the transmission speed, in order to balance the encoding speed and quality. Then the predictive video block is formed during motion compensation using the best motion vector. As described above, the residual video block is formed by subtracting the predicted video block from the original video block. The transform is then applied to the residual block, and the transform coefficients are quantized and entropy encoded to further reduce the bit rate.

On the other hand, figa-4D show integer-pixel samples (also called full-pixel) in solid rectangles with capital letters. For any given integer-pixel sampling, a total of 15 sub-pixel positions are provided, which are shown for the integer-pixel sampling “C3” and marked a-o in FIGS. 4A-4D. In H.264 / AVC, the half-pixel positions “b”, “h” and “j” can be calculated first using a one-dimensional Wiener 6-tap filter. As noted, the half-pixel position “j” may require two levels of this interpolation using 6-tap filtering. In this case, for the pixel position “j”, the interpolation filter is applied first in the horizontal direction and then in the vertical direction, possibly with an intermediate rounding of the values of the interpolated pixel after interpolation in one dimension to ensure that the data can be stored in sixteen-bit storage elements. The remaining quarter-pixel positions are then interpolated through filtering, which uses a bilinear filter and already calculated half-pixel samples.

The ITU-T SG16 / Q.6 / VCEG (video coding expert group) is investigating coding technologies that offer higher coding efficiency than H.264 / AVC. This study is being conducted in the KTA Forum (Key Technology Areas). One of the coding tools that are adapted in the KTA is called adaptive interpolation filter (AIF). AIF offers greater coding efficiency than H.264 / AVC, in particular for high-resolution video sequences (e.g. 720p). In AIF, an interpolation filter for each sub-pixel position is analytically calculated for each video frame by minimizing the energy of prediction errors. Analytically extracted adaptive filtering coefficients for each frame are then predicted, quantized, encoded, and sent in the video bit stream.

The technologies of this disclosure can achieve comparable coding improvements with achievable AIFs while reducing implementation complexity relative to AIFs. The described technologies may not require an analytical process in order to extract filtering coefficients for each video frame. Instead, in one example, various sets of filter coefficients are pre-computed and pre-defined for a real-time application. In this case, pre-setting the filter coefficients can be based on the AIF applied to the test sequences or other selection criteria.

Three different types of AIF circuits are fitted in the KTA. The first is a two-dimensional non-separable AIF (NS-AIF), the second is a separable AIF (S-AIF), and the third is an AIF with directional filters (DAIF). However, all three AIF schemes use the same analytical process to extract filtering coefficients, which is explained below using an inseparable AIF as an example.

, где i, j=0,..., 5, и SP представляет одну из 15 субпиксельных позиций (a-o), показанных на фиг.4A-4D. Следует отметить, что 6 из 15 субпиксельных позиций "a", "b", "c", "d", "h" и "l" являются одномерными субпиксельными позициями, которые могут использовать 6-отводный фильтр для интерполяции. Кроме того, допустим, что прогнозирующие пикселы в целопиксельных позициях (A1-F6 на фиг.3) в опорном кадре принимают пиксельные значения

, где i, j=0,..., 5. Таким образом, A1 принимает значение

,..., A6 принимает значение

,..., F1 принимает значение

,..., и F6 принимает значение

. Затем интерполированное значение

в субпиксельной позиции SP,

, может вычисляться с использованием следующего уравнения:Suppose a two-dimensional inseparable 6x6 filter has coefficients

where i, j = 0, ..., 5, and SP represents one of the 15 sub-pixel positions (ao) shown in FIGS. 4A-4D. It should be noted that 6 of the 15 sub-pixel positions "a", "b", "c", "d", "h" and "l" are one-dimensional sub-pixel positions that can use a 6-tap filter for interpolation. Furthermore, suppose that the predictive pixels at integer-pixel positions (A1-F6 in FIG. 3) in the reference frame take pixel values

, where i, j = 0, ..., 5. Thus, A1 takes the value

, ..., A6 takes the value

, ..., F1 takes on the value

, ..., and F6 takes the value

. Then the interpolated value

in subpixel position SP,

can be calculated using the following equation:

является пиксельным значением в текущем видеокадре в позиции

.Let be

is the pixel value in the current video frame at position

.

,

,

- это вектор движения,

- это целочисленный компонент вектора движения, и

- это смещение фильтра. Например, в случае 6-отводного фильтра,

. Для каждой субпиксельной позиции SP энергия

ошибок прогнозирования между фактическим пиксельным значением в текущем кадре и интерполированным значением может накапливаться для всех пикселов, которые имеют точность вектора движения, соответствующую субпиксельной позиции SP. Энергия

ошибок прогнозирования вычисляется с использованием следующего уравнения:Where

is a motion vector,

is the integer component of the motion vector, and

is the filter offset. For example, in the case of a 6-tap filter,

. For each subpixel position SP energy

prediction errors between the actual pixel value in the current frame and the interpolated value can accumulate for all pixels that have a motion vector accuracy corresponding to the sub-pixel position SP. Energy

prediction errors is calculated using the following equation:

(2)

(2)

относительно коэффициентов

фильтрации. Число уравнений в этом случае равно числу коэффициентов фильтрации, используемых для текущей субпиксельной позиции

. Для каждой двумерной субпиксельной позиции "e", "f", "g", "i", "j", "k", "m", "n" и "o" с использованием 6x6-отводного двумерного фильтра может разрешаться система из 36 уравнений с 36 неизвестными. Для оставшихся субпиксельных позиций "a", "b", "c", "d", "h" и "l", которые могут требовать только одномерного фильтра, системы из 6 уравнений могут разрешаться.For each of the subpixel positions ao, a separate system of equations can be established by calculating the derivative

relative to the coefficients

filtering. The number of equations in this case is equal to the number of filtering coefficients used for the current subpixel position

. For each two-dimensional subpixel position "e", "f", "g", "i", "j", "k", "m", "n" and "o" using a 6x6-tap two-dimensional filter, a system can be resolved of 36 equations with 36 unknowns. For the remaining subpixel positions "a", "b", "c", "d", "h" and "l", which may require only a one-dimensional filter, systems of 6 equations can be resolved.

(3)

(3)

To summarize the above, an exemplary process for extracting and applying AIF filters can have the following steps:

движения могут быть оценены для каждого видеокадра, который должен кодироваться. Во время оценки движения неперестраиваемый интерполяционный фильтр (к примеру, интерполяционный процесс H.264/AVC) может применяться.1. Vectors

Motion can be estimated for each video frame to be encoded. During motion estimation, a non-tunable interpolation filter (e.g., H.264 / AVC interpolation process) may be applied.

может накапливаться по текущему видеокадру. Затем коэффициенты

адаптивной фильтрации могут вычисляться для каждой субпиксельной позиции

независимо посредством минимизации энергии ошибок прогнозирования, как указано в двух вышеприведенных уравнениях энергии прогнозирования.2. Using these motion vectors, the energy of prediction errors for each subpixel position

may accumulate on the current video frame. Then the odds

adaptive filtering can be calculated for each subpixel position

independently by minimizing the energy of prediction errors, as indicated in the two prediction energy equations above.

3. New motion vectors can then be estimated. During the motion estimation process, adaptive interpolation filters calculated in step 2 can be applied. Using adaptive interpolation filters, the motion estimation errors caused by spectral overlapping, camera noise, etc. are reduced, and better motion prediction can be achieved.

All three existing AIF schemes use an analytical process identical to the process indicated above. The differences between these schemes are mainly in the number of unique filtering coefficients used, in that the interpolation filters are separable or inseparable, and in what filter support (i.e. integer-pixel positions used for interpolation) are used. In each of these schemes, certain symmetry restrictions for AIF filters are imposed to reduce the number of filtering coefficients that must be sent in the video bitstream.

As mentioned above, in H.264 / AVC, interpolation is performed using intermediate rounding operations. Traditionally, half-pixel positions ("b", "h", and "j") are filtered and rounded before they are used to obtain quarter-pixel positions. In addition, during bilinear filtering, which is used to obtain quarter-pixel values, rounding up can be used (i.e. (a + b + 1) >> 1 can be used in traditional bilinear filters). Such frequent and biased rounding operations can reduce the accuracy of interpolation filters. In particular, simulation shows that interpolation filters with higher accuracy, extracted directly from the Wiener 6-tap filter, and the bilinear filter used in H.264 / AVC, provide significant efficiency compared to H.264 / AVC, in particular for high definition video sequences such as 720p sequences. A significant part of the efficiency achieved through AIF schemes (NS-AIF, S-AIF and DAIF) may come from the fact that these AIF schemes use filters with higher accuracy than H.264 / AVC.

According to this disclosure, a relatively simple switched offset interpolation filter (SIFO) scheme may be used. Instead of adaptively extracting filtering coefficients for each frame, which is a very costly process on the encoder side, the SIFO scheme selects sets of non-tunable filters, for example, that can be set accordingly by means of many different predefined sets of interpolation filtering coefficients. The selection of the interpolation filter can be performed either at the frame level or at the sub-pixel position level, which may be referred to as frame-based selection or selection based on samples, respectively. If desired, DC offsets can also be added after prediction, and DC offsets can also be added based on frames or samples. Thus, one or more DC offsets can be set for each frame, different DC offsets (or sets of DC offsets) can be set for different video blocks of the frame. Different DC offsets can be set for each pixel and subpixel interpolated position. Frames are one example of coded units, and the phrase “frame-based selection” is more commonly referred to as coded-unit selection. In some cases, a series of consecutive macroblocks or parts of a frame, for example, can be encoded separately as separate coded units.

Various sets of non-tunable filters can be used by the motion compensation modules 35 and 55 of FIGS. 2 and 3, respectively. Examples include a standard H.264 filter set, a filter set based on H.264 / AVC, but with higher accuracy (without intermediate rounding for half-pixel positions and offset rounding for quarter-pixel positions), or one or more sets of custom interpolation filters. A set or sets of custom interpolation filters can be obtained by using a set of training video sequences. Adaptive filters that minimize the energy of prediction errors can be extracted by applying the analytic process used in AIF circuits to training video set. Then, the average filters for each subpixel position can be calculated and used as custom interpolation filters, which are applied by the motion compensation units 35 and 55, as described herein.

The average filter for a particular sub-pixel position can also be normalized so that all filter coefficients add up to a total of 1.0. The middle filter can also be quantized with a certain precision with a fixed point (for example, 13-bit precision). The use of fixed-point precision in the filter coefficient ensures that implementations for different platforms do not have deviations. In addition, a true fixed-point implementation may be derived from such predetermined filtering coefficients. Custom filters can have different characteristics. In particular, different filters can be inseparable filters or separable filters, and different filters can specify different filter supports (for example, 6x6 or 4x4).

In any case, the filter set that offers the best prediction in any given example (for example, the lowest energy of prediction errors) can be selected by the motion compensation module 35 during the encoding process. On the other hand, the selection of the filter set may be a value based on samples or based on frames, which means that filters can be selected for each frame (or other coded unit) or can be selected for different video blocks of each frame. When multiple reference frames are used, different filter sets can be selected for different reference frames. In addition, some reference frames may have a filter selection based on frames, while other reference frames may have a filter selection based on samples. When frame-based filter selection is used, and when a custom filter set is selected, standard filters can still be applied to half-pixel positions (for example, b, h and j positions), while a custom filter set can be applied for other subpixel positions.

An offset can also be added during interpolation. Adding DC offsets to prediction values can help capture lighting changes between different video frames, such as flashes, twilight sky, etc. In the schema of this disclosure, DC offset can be applied at the frame level (all pixels in the frame use the same DC offset) or at the sample position level (different pixel locations have different DC offsets). When sample-based DC offsets are used, a total of sixteen DC offsets may need to be sent in the video bit stream so that integer-pixel locations and fifteen possible non-integer-pixel locations can specify different offsets. Sample-based DC bias can be useful because it can act as a simple but effective tool for motion segmentation. Consider, for example, a video frame that contains a darkening sky as a fixed background and a moving object as a foreground. In this case, the background and foreground can have different degrees of lighting changes, and using sample-based DC offsets can help capture different degrees of lighting changes in the foreground and background.

When multiple reference frames are used, different DC offsets can be calculated and sent for different reference frames. In addition, some reference frames may have frame-based DC bias, while others may have sample-based DC biases. One particular scheme is to use sample-based DC offsets only in the reference frame that is the closest in time to the current frame, and use frame-based DC offset for all other reference frames. Information can be encoded in a bit stream to tell the decoder the type of offsets used and the magnitude of such offsets at various pixel locations.

A subpixel motion vector is referred to as a subpixel position in the reference image to be interpolated. H.264 defines one interpolation process for subpixels in which subpixels b and h (see FIGS. 4A-4D) can be calculated by horizontal and vertical filtering using a 6-tap filter having tap values (1, -5, 20, 20, -5, 1), as follows:

b1 = C1-5 * C2 + 20 * C3 + 20 * C4-5 * C5 + C6

where "C1", "C2", "C3", "C4", "C5" and "C6" represent the six nearest integer pixels that surround "b" in the horizontal direction, and the pixels "C3" and "C4" are closest, “C2” and “C5” are the next closest, and “C1” and “C6” are the next closest.

h1 = A3-5 * B3 + 20 * C3 + 20 * D3-5 * E3 + F3

where "A3", "B3", "C3", "D3", "E3" and "F3" represent the six nearest integer pixels that surround the "h" in the vertical direction, the pixels "C3" and "D3" being the nearest , "B3" and "E3" are the next closest, and "A3" and "F3" are the next closest.

In this case, the values of "b" and "h" can be set as follows:

b = max (0, min (255, (b1 + 16) >> 5))

h = max (0, min (255, (h1 + 16) >> 5))

where ">>" is the shift operation to the right. In this disclosure, ">>" represents a shift operation to the right, and "<<" represents a shift operation to the left.

To interpolate the subpixel "j", the intermediate value "j1" is first extracted as follows:

j1 = aa1-5 * bb1 + 20 * b1 + 20 * hh1-5 * ii1 + jj1,

where the intermediate values, designated as "aa1", "bb1", "hh1", "ii1" and "jj1", are retrieved by applying a 6-tap filter horizontally in a manner identical to calculating b1 at the positions "aa", "bb", " hh "," ii "and" jj ". The final value of j can be calculated as follows:

j = max (0, min (255, (j1 + 512) >> 10)).

Pixels in quarter-pixel positions labeled "a", "c", "d", "l", "f", "i", "k" and "n" can be extracted according to table 1 by averaging the two nearest pixels in integer- and half-pixel positions and the subsequent application of rounding up.

Table 1 a = (C3 + b + 1) >> 1 c = (C4 + b + 1) >> 1 d = (C3 + h + 1) >> 1 l = (D3 + h + 1) >> 1 f = (j + b + 1) >> 1 i = (j + h + 1) >> 1 k = (j + ee + 1) >> 1 n = (j + hh + 1) >> 1

Pixels at quarter-pixel positions labeled “e”, “g”, “m”, and “o” are calculated according to Table 2 by averaging the two nearest pixels at half-pixel positions in the diagonal direction and then applying rounding up.

table 2 e = (b + h + 1) >> 1 g = (b + ee + 1) >> 1 m = (h + hh + 1) >> 1 o = (ee + hh + 1) >> 1

The final values of these quarters of pixels can be finally cut off to values in the interval (0.255).

In order to maintain the highest possible accuracy in the intermediate stages, all shear, rounding and clipping operations can be eliminated until the very last stage of the interpolation process. In contrast, in an implementation defined according to the H.264 standard, pixels at quarter-pixel positions are labeled “a”, “c”, “d”, “l”, “f”, “i”, “k” and “n” are extracted according to table 3 by averaging the two nearest pixels at integer and half-pixel positions “b”, “h”, “j”, “ee” and “hh”, with rounding up.

Table 3 a = (C3 << 5 + b1 + 32) >> 6 c = (C4 << 5 + b1 + 32) >> 6 d = (C3 << 5 + h1 + 32) >> 6 l = (D3 << 5 + h1 + 32) >> 6 f = (j1 >> 5 + b1 + 32) >> 6 i = (j1 >> 5 + h1 + 32) >> 6 k = (j1 >> 5 + ee1 + 32) >> 6 n = (j1 >> 5 + hh1 + 32) >> 6

In this case, the intermediate values designated as “ee1” and “hh1” are extracted by applying the 6-tap filter vertically and horizontally in the same manner as in the calculation of “h1” and “b1” mentioned above, but at the positions “ee” and hh. Rounding may not be allowed at this stage in the formation of “ee1” and “hh1” as described herein.

Pixels in quarter-pixel positions labeled “e”, “g”, “m”, and “o” can be calculated according to Table 4 by averaging the two nearest pixels at half-pixel positions in a diagonal direction with rounding up.

Table 4 e = (b1 + h1 + 32) >> 6 g = (b1 + ee1 + 32) >> 6 m = (h1 + hh1 + 32) >> 6 o = (ee1 + hh1 + 32) >> 6

The final values of these quarters of pixels can be clipped to values in the range (0.255). By maintaining the full accuracy of the intermediate values, the interpolated subpixels should be more accurate. In particular, half-pixel values at locations “b,” “h,” “ee,” and “hh” may be unrounded values designated as “b1,” “h1,” “ee1,” and “hh1” for quarter-pixel generation.

Unfortunately, interpolating the value of "j" can be more complicated, since interpolating at position j requires two levels of half-pixel interpolation. In particular, to interpolate "j", twenty bits may be required to represent an intermediate value of "j1". In this case, the intermediate version “j1” for “j” may be rounded to correspond to sixteen bits. In this case, the operations explained above can be implemented with a slight modification so that all operations can correspond to sixteen-bit data structures without compromising accuracy. Interpolation "b" and "h" may be identical to the interpolation specified above, except that their intermediate versions "b1" and "h1" can be saved for another subpixel interpolation. The following tables show the interpolation process for other subpixels in sixteen-bit storage elements. In the tables below, the operations specified in each column are performed sequentially for the corresponding table. Shaded lines contain operations that are performed at the end of the process and are not used to interpolate other sub-pixel positions.

Table 5
positions {a, c, d, l} of FIGS. 4A-4D Operation Comment Minimum value Maximum value Register size r1 = x r1 - integer pixel x 0 255 8u r1 = r1 << 5 r1 is 32 * x 0 8160 13u r2 = y0 r2 is y0 (y0 is one-dimensional -2550 10710 15s half-pixel, such as b1, h1, ee1 and hh1, before shifting towards lower bits) r1 = r1 + r2 r1 is 32 * x + y0 -2550 18870 16s r1 = r1 + 32 r1 is 32 * x + y0 + 32 -2518 18902 16s r1 = r1 >> 6 r1 is (32 * x + y0 + 32) >> 6 -39 295 11s r1 = max (0, r1) clipping r1 on the low side 0 295 10u r1 = min (255, r1) clipping r1 on the high side 0 255 8u

Table 6
positions {e, g, m, o} of FIGS. 4A-4D Operation Comment Minimum value Maximum value Register size r1 = y0 r1 is y0
(y0 is a one-dimensional half-pixel, such as b1, h1, ee1 and hh1, before shifting towards the lower bits) -2550 10710 15s r2 = y1 r2 is y1 (y1 is one-dimensional -2550 10710 15s half-pixel, such as b1, h1, ee1 and hh1, before shifting towards lower bits) r1 = r1 + r2 r1 is y0 + y1 -5100 21420 16s r1 = r1 + 32 r1 is y0 + y1 + 32 -5068 21452 16s r1 = r1 >> 6 r1 is (y0 + y1 + 32) >> 6 -79 335 11s r1 = max (0, r1) clipping r1 on the low side 0 335 10u r1 = min (255, r1) clipping r1 on the high side 0 255 8u

Table 7 below illustrates a sixteen-bit implementation of extracting a pixel value at the pixel location “j”. All one-dimensional half-pixels "aa1", "bb1", "b1", "hh1", "ii1" and "jj1" are obtained without any shift towards the lower bits. The cut-off range, to prevent overflow, is (-2372.18640). In addition, the hatched lines in the table indicate operations that are performed only at the end of the interpolation process, and these results are not used to interpolate other sub-pixel positions. The intermediate value "j1" must be stored for interpolation of "f", "i", "k" and "n".

Table 8 below shows the steps that can be performed for a sixteen-bit interpolation implementation {f, i, k, n}, which are positions that are used to interpolate the intermediate value "j1".

Table 8
positions {f, i, k, n} of FIGS. 4A-4D Operation Comment Minimum value Maximum value Register size r1 = y0 r1 is y0 (a one-dimensional half-pixel, such as b1, h1, ee1 and hh1, before shifting towards the lower bits) -2550 10710 15s r2 = j1 r2 is j1 (two-dimensional half-pixel j1 before shifting towards the lower bits) -9914 26232 16s r2 = r2 >> 1 r2 is j1 >> 1 -4957 13116 15s r1 = r1 + r2 r1 is y0 + (j1 >> 1) -7507 23826 16s r1 = r1 + 32 r1 is y0 + (j1 >> 1) +32 -7491 23842 16s r1 = r1 >> 6 r1 is (y0 + (j1 >> 1) +32) >> 6 -235 745 11s r1 = max (0, r1) clipping r1 on the low side 0 745 10u r1 = min (255, r1) clipping r1 on the high side 0 255 8u

Briefly, the explanation above shows a complete technology for interpolating all possible subpixel locations "a", "b", "c", "d", "e", "f", "g", "h", "i", "j "," k "," l "," m "," n "and" o "without exceeding sixteen-bit storage elements, which is desirable for implementation. Most intermediate half-pixel values remain unrounded for quarter-pixel interpolation purposes. Nevertheless, the pixel location “j” is a special case, which may require rounding of the intermediate results in order to form its half-pixel value, since the pixel location “j” requires two levels of half-pixel interpolation. In accordance with this disclosure, however, rounding at the pixel location “j” does not affect the accuracy of the final value of the quarters of the pixels that are based on “j”.

5 is a flowchart consistent with one or more aspects of video encoding in accordance with this disclosure. Figure 5 is explained from the point of view of the video encoder 50 shown in figure 2. In particular, the motion compensation module 35 generates predictive data (101), which includes interpolation based on a plurality of predetermined interpolation filters. Motion compensation module 35 then encodes video data, for example, a video block based on predictive data associated with each of a predetermined interpolation filter (102). For example, motion compensation module 35 may activate an adder 48 to form residual blocks associated with each of a predetermined interpolation filter so that various residual blocks can be compared with each other to find the best residual block. Motion compensation module 35 selects video data that reaches the highest compression levels (103). In particular, the selected video data is illustrated as outputting a “residual block” of adder 48, which is the final result. This final result of the adder 48 is converted by the transform module 38, quantized by the quantization module 40, and entropy encoded by the entropy encoding module 46.

Motion compensation module 35 also identifies a particular interpolation filter that is associated with selected residual video data that reaches the highest compression levels (104). Motion compensation unit 35 redirects the interpolation syntax element to entropy encoding unit 46. The interpolation syntax element identifies the particular interpolation filter that is used. Entropy encoding unit 46 encodes a syntax element (105) to form part of the output bitstream. Thus, the interpolation filter used in the encoder 50 (for example, the filter coefficients used or other filter schemes or parameters) can be transmitted to the target device so that the target device can perform proper interpolation filtering during video recovery.

Interpolation filtering may be set based on each frame or may be set based on each video block. In still other cases, interpolation filtering may be specified based on each sample, or may be defined based on each subsample. The term “video units” may be referred to as video frames, video blocks, video samples or video samples. In any case, video encoder 50 may encode various syntax elements to identify various interpolation filters for various reference video units, which are used to generate predictive data for various encoded units of video data. Alternatively, the video encoder may encode various syntax elements to identify various interpolation filters that are used for various video blocks of the reference video data. In any case, interpolation filters can be predefined using test video sequences. In this case, for example, adaptive interpolation filtering (AIF) can be applied to video sequences to specify various interpolation filters that achieve the desired results, and these interpolation filters can be programmed in the motion compensation unit 35.

6 is a flowchart consistent with one or more aspects of video decoding in accordance with this disclosure. 6 is explained from the point of view of the video encoder 50 shown in figure 2. In particular, video encoder 60 receives encoded video data (111) and receives a syntax element that identifies an interpolation filter from a plurality of interpolation filters (112). The encoded bitstream may include both encoded residual video data and a syntax element that identifies the interpolation filter that is used in the encoder. Entropy decoding unit 52 may entropy decode the received bitstream to parse the converted and quantized residual blocks that are sent to inverse transform quantization unit 56 and inverse transform unit 58, and syntactically analyze the interpolation syntax element and motion vectors that are sent to unit 55 motion compensation.

Motion compensation module 55 generates predictive data (113) that includes interpolations based on an interpolation filter identified by a syntax element. Video decoder 60 decodes the video data based on the predictive data (114). In particular, the predictive data generated by the motion compensation unit 55 may be combined with the residual data through an adder 64 to recover video data in the pixel domain. The decoded video can then be stored in the reference frame storage unit 62 for later use in prediction and can also be output for presentation to the user, for example, through a display or other output device.

As in the case of encoding, interpolation filtering during the decoding process can be set based on each frame or can be set based on each video block, each subsample, or each sample. In particular, video decoder 60 may decode various syntax elements to identify different interpolation filters for various reference video units, which are used to generate predictive data for various encoded video units, or alternatively, video decoder 60 may decode various syntax elements to identify different interpolation filters that are used for various video blocks of the reference video data. In any case, the interpolation filters can be predefined using test video sequences, and these interpolation filters can be programmed in the motion compensation module 55 so that the syntax can be used to identify and activate the appropriate interpolation filter within the motion compensation module 55.

7 illustrates interpolation technology that can be performed by motion compensation module 35 during the encoding process or by motion compensation module 55 during the decoding process. As shown in FIG. 7, the motion compensation module 35 or 55 generates half-pixel values based on integer-pixel values (121). For example, these half-pixel values may correspond to pixel values at locations “b” and “h” of FIGS. 4A-4D. Motion compensation unit 35 or 55 rounds these half-pixel values to form half-pixel interpolated values (122), and also stores unrounded versions of half-pixel values (123). Motion compensation modules 35 and 55 may have memory elements to facilitate this temporary storage of intermediate values corresponding to non-rounded versions of half-pixel values. Motion compensation module 35 or 55 can then generate quarter-pixel values based on non-rounded half-pixel values and integer-pixel values (124). These quarter-pixel values, for example, can correspond to the locations "a", "c", "d" and "l" shown in figa-4D.

As explained above, half-pixel values at location “j” in FIGS. 4A-4D may require some rounding of the intermediate values, since location “j” may require two levels of half-pixel interpolations. FIG. 8 illustrates interpolation technology that can be performed by motion compensation module 35 during the encoding process or by motion compensation module 55 during the decoding process to generate all interpolated values (ao shown in FIGS. 4A-4D) as part of an implementation with a fixed number of bits, for example, for sixteen bits.

Similar to the process of FIG. 7, the process of FIG. 8 begins by generating, with the module 35 or 55, motion compensation of half-pixel values based on integer-pixel values (131), for example, which can correspond to pixel values at locations “b” and “h” of FIG. .4A-4D. Motion compensation unit 35 or 55 then rounds these half-pixel values to form half-pixel interpolated values (132), and also stores unrounded versions of half-pixel values (133).

Further, the motion compensation unit 35 or 55 generates an additional half-pixel value (for example, a value corresponding to the location “j” in FIGS. 4A-4D) based on the non-rounded half-pixel values. Motion compensation module 35 or 55 rounds an additional half-pixel value (for example, the value corresponding to the location “j” in FIGS. 4A-4D), if necessary, to ensure that this value is within a fixed bit depth (for example, in sixteen bits) (135) without compromising the accuracy of the quarter pixels based on the "j". Motion compensation module 35 or 55 can then generate quarter-pixel values based on non-rounded half-pixel values, additional half-pixel values, and integer-pixel values (136). Thus, all quarter-pixel values shown by shading in FIG. 4B (for example, locations “a”, “c”, “d”, “e”, “f”, “g”, “i”, “k "," l "," m "," n "and" o "may be formed). The above tables and explanation provide more specific details regarding the operations that can be performed in this process.

The techniques of FIGS. 7 and 8 can be applied to many different interpolation contexts. In particular, these technologies may be used with or without other technologies described herein that use predefined interpolation filters. In other words, the technologies of FIGS. 7 and 8, which eliminate or significantly reduce intermediate rounding, can be used during interpolation according to one or predefined interpolation filters, but can also be used for interpolation that uses a non-tunable interpolation filter, or, possibly, with technologies adaptive refinement interpolations. Accordingly, the various technologies described herein do not have to be performed together, but may have the desired application with many different settings or in many cases of interpolation.

The technologies of this disclosure can be implemented in a wide range of devices or devices, including a cordless handset and integrated circuit (IC) or IC set (i.e., a chipset). All described components, modules or units are provided in order to emphasize functional aspects, and do not necessarily require implementation through various hardware modules, etc.

The technologies described herein may be implemented in hardware, software, firmware, or any combination thereof. If implemented in hardware, any features described as modules, nodes, or components can be implemented together in an integrated logic device or separately as discrete but interacting logic devices. If implemented in software, technologies can be implemented at least in part by a computer-readable medium containing instructions that, when executed on a processor, perform one or more of the methods described above, the computer-readable medium may comprise a computer-readable storage medium, and may be part of a computer software product, which may include packaging. Computer-readable storage media may include random access memory (RAM), such as synchronous dynamic random access memory (SDRAM), read-only memory (ROM), non-volatile random access memory (NVRAM), electrically erasable programmable read-only memory (EEPROM), flash - memory, magnetic or optical storage media, etc. Additionally or alternatively, the technologies can be implemented, at least in part, through a computer-readable communication medium that transfers or transmits code in the form of instructions or data structures, and which can be accessed, read, or executed by a computer.

The code can be executed by one or more processors, for example, one or more digital signal processors (DSPs), general purpose microprocessors, specialized integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other equivalent integrated or discrete logic circuits. Accordingly, the term "processor" as used herein may mean any of the above structure or other structure suitable for implementing the technologies described herein. In addition, in some aspects, the functionality described herein may be provided within specialized software modules or hardware modules configured to encode or decode or integrated into a combined video encoder-decoder (codec). In addition, technologies can be fully implemented in one or more circuits or logic elements.

Various aspects of the disclosure of the essence are described. These and other aspects are within the scope of the attached claims.

Claims (26)

1. A method for interpolating predictive video data for encoding video, the method comprising the steps of: generating a plurality of half-pixel values based on a plurality of integer-pixel values of the video data; round a plurality of half-pixel values to form a plurality of half-pixel interpolated values; save a plurality of half-pixel values as a plurality of non-rounded versions of a plurality of half-pixel values; generating a plurality of quarter pixel values based on a plurality of unrounded versions of a plurality of half-pixel values and a plurality of integer-pixel values; generating predictive data using a plurality of half-pixel interpolated values and a plurality of quarter-pixel values; and store predictive data for use in video encoding or video decoding. 2. The method according to claim 1, additionally containing stages in which: form another half-pixel value in addition to the set of half-pixel values based on the set of unrounded versions of the set of half-pixel values; and generating a plurality of additional quarter-pixel values in addition to a plurality of quarter-pixel values based on another half-pixel value and a plurality of non-rounded versions of the plurality of half-pixel values. 3. The method according to claim 1, additionally containing stages in which: a plurality of half-pixel interpolated values and a plurality of quarter-pixel values are stored in a sixteen-bit data structure; generating another half-pixel value in addition to the plurality of half-pixel values based on the plurality of unrounded versions of the plurality of half-pixel values; round another half pixel value so that another half pixel value can be stored in a sixteen-bit data structure; and generating a plurality of additional quarter-pixel values in addition to a plurality of quarter-pixel values based on another half-pixel value and a plurality of non-rounded versions of the plurality of half-pixel values. 4. The method according to claim 1, further comprising the step of encoding one or more video blocks using predictive data. 5. The method according to claim 1, further comprising the step of decoding one or more video blocks using predictive data. 6. A device that interpolates predictive video data for encoding video, wherein the device includes a motion compensation module that: generates a plurality of half-pixel values based on a plurality of integer-pixel values of the video data; rounds multiple half-pixel values to form multiple half-pixel interpolated values; saves multiple half-pixel values as multiple non-rounded versions of multiple half-pixel values; forms a plurality of quarter pixel values based on a plurality of unrounded versions of a plurality of half-pixel values and a plurality of integer-pixel values; generates predictive data using multiple half-pixel interpolated values and multiple quarter-pixel values; and stores predictive data for use in video encoding or video decoding. 7. The device according to claim 6, in which the motion compensation module: generates another half-pixel value in addition to a plurality of half-pixel values based on a plurality of unrounded versions of the plurality of half-pixel values; and forms a plurality of additional quarter-pixel values in addition to a plurality of quarter-pixel values based on another half-pixel value and a plurality of non-rounded versions of the plurality of half-pixel values. 8. The device according to claim 6, in which the motion compensation module: stores a plurality of half-pixel interpolated values and a plurality of quarter-pixel values in a sixteen-bit data structure; generates another half-pixel value in addition to a plurality of half-pixel values, based on a plurality of unrounded versions of the plurality of half-pixel values; rounds another half-pixel value so that another half-pixel value can be stored in a sixteen-bit data structure; and forms a plurality of additional quarter-pixel values in addition to a plurality of quarter-pixel values based on another half-pixel value and a plurality of non-rounded versions of the plurality of half-pixel values. 9. The device according to claim 6, wherein the device comprises a video encoder that includes a motion compensation module, wherein the video encoder encodes one or more video blocks with prediction using predictive data. 10. The device according to claim 6, wherein the device comprises a video decoder that includes a motion compensation module, wherein the video decoder predicts one or more video blocks using predictive data. 11. The device according to claim 6, wherein the device comprises an integrated circuit. 12. The device according to claim 6, wherein the device comprises a microprocessor. 13. A computer-readable storage medium containing instructions that, when executed, instruct the device to interpolate predictive video data for encoding video, the instructions instruct the device: generate a plurality of half-pixel values based on a plurality of integer-pixel values of the video data; round off multiple half-pixel values to form multiple half-pixel interpolated values; save multiple half-pixel values as multiple unrounded versions of multiple half-pixel values; generate a plurality of quarter pixel values based on a plurality of unrounded versions of a plurality of half-pixel values and a plurality of integer-pixel values; generate predictive data using multiple half-pixel interpolated values and multiple quarter-pixel values; and save predictive data for use in video encoding or video decoding. 14. The computer-readable storage medium according to claim 13, further comprising instructions that, when executed, instruct the device to: generate another half-pixel value in addition to a plurality of half-pixel values based on a plurality of non-rounded versions of the plurality of half-pixel values; and generate a plurality of additional quarter-pixel values in addition to a plurality of quarter-pixel values based on another half-pixel value and a plurality of non-rounded versions of the plurality of half-pixel values. 15. The computer-readable storage medium according to claim 13, further comprising instructions that, when executed, instruct the device: to store a plurality of half-pixel interpolated values and a plurality of quarter-pixel values in a sixteen-bit data structure; generate another half-pixel value in addition to the plurality of half-pixel values based on the plurality of unrounded versions of the plurality of half-pixel values; Round another half-pixel value so that another half-pixel value can be stored in a sixteen-bit data structure; and generate a plurality of additional quarter-pixel values in addition to a plurality of quarter-pixel values based on another half-pixel value and a plurality of non-rounded versions of the plurality of half-pixel values. 16. The computer-readable storage medium according to item 13, further comprising instructions that, when executed, instruct the device: encode one or more video blocks using predictive data. 17. The computer-readable storage medium according to item 13, further comprising instructions that, when executed, instruct the device to: decode one or more video blocks using predictive data. 18. A device that interpolates predictive video data for encoding video, the device comprising: means for generating a plurality of half-pixel values based on a plurality of integer-pixel values of video data; means for rounding a plurality of half-pixel values to form a plurality of half-pixel interpolated values; means for storing a plurality of half-pixel values as a plurality of non-rounded versions of a plurality of half-pixel values; means for generating a plurality of quarter pixel values based on a plurality of unrounded versions of a plurality of half-pixel values and a plurality of integer-pixel values; means for generating predictive data using a plurality of half-pixel interpolated values and a plurality of quarter-pixel values; and means for storing predictive data for use in video encoding or video decoding. 19. The apparatus of claim 18, further comprising: means for generating another half-pixel value in addition to a plurality of half-pixel values based on a plurality of non-rounded versions of the plurality of half-pixel values; and means for generating a plurality of additional quarter-pixel values in addition to a plurality of quarter-pixel values based on another half-pixel value and a plurality of non-rounded versions of the plurality of half-pixel values. 20. The apparatus of claim 18, further comprising: means for storing a plurality of half-pixel interpolated values and a plurality of quarter-pixel values in a sixteen-bit data structure; means for generating another half-pixel value in addition to the plurality of half-pixel values based on the plurality of unrounded versions of the plurality of half-pixel values; means for rounding another half-pixel value so that another half-pixel value can be stored in a sixteen-bit data structure; and means for generating a plurality of additional quarter-pixel values in addition to a plurality of quarter-pixel values based on another half-pixel value and a plurality of non-rounded versions of the plurality of half-pixel values. 21. The device according to p, optionally containing: means for encoding one or more video blocks using predictive data. 22. The device according to p, further comprising: means for decoding one or more video blocks using predictive data. 23. A device that encodes video data, the device comprising: a video encoder including a motion compensation module that generates a plurality of half-pixel values based on a plurality of integer-pixel values of the reference video data, rounds a plurality of half-pixel values to form a plurality of half-pixel interpolated values, stores a plurality of half-pixel values as a set of non-rounded versions of a set of half-pixel values, forms a set of quarter-pixel values on based on a plurality of unrounded versions of a plurality of half-pixel values and a plurality of integer-pixel values, generates predictive data using a plurality of half-pixel interpolated values and a plurality of quarter-pixel values and encodes video data using the predictive data; and a wireless transmitter that transmits video encoded by the motion compensation module to another device. 24. The device according to item 23, wherein the device comprises a portable wireless telephone. 25. A device that decodes video data, the device comprising: a wireless receiving device that receives video data; and a video decoder including a motion compensation module that decodes the video data, and when decoding the video data, the motion compensation module: generates a plurality of half-pixel values based on the plurality of integer-pixel reference video data; rounds multiple half-pixel values to form multiple half-pixel interpolated values; saves multiple half-pixel values as multiple non-rounded versions of multiple half-pixel values; forms a plurality of quarter pixel values based on a plurality of unrounded versions of a plurality of half-pixel values and a plurality of integer-pixel values; generates predictive data using multiple half-pixel interpolated values and multiple quarter-pixel values; and decodes video data using predictive data. 26. The device according A.25, while the device contains a portable telephone wireless communication.

Images