EP4315170A1 - Autoencodeur multimodal a fusion de donnees latente amelioree - Google Patents
Autoencodeur multimodal a fusion de donnees latente amelioreeInfo
- Publication number
- EP4315170A1 EP4315170A1 EP22719247.3A EP22719247A EP4315170A1 EP 4315170 A1 EP4315170 A1 EP 4315170A1 EP 22719247 A EP22719247 A EP 22719247A EP 4315170 A1 EP4315170 A1 EP 4315170A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- latent
- vector
- autoencoder
- modality
- encoders
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G06N3/0455—Auto-encoder networks; Encoder-decoder networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0475—Generative networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Definitions
- the present invention relates to the field of machine learning. More specifically, the invention relates to the field of autoencoders.
- Autoencoders are a recent technique for transforming a complex data source into a high-level representation.
- Autoencoders are a type of artificial neural networks trained to perform efficient data encoding in an unsupervised manner.
- An autoencoder consists of a first neural network, which encodes an input vector generally denoted x into a compressed vector (also called latent vector) generally denoted z, and a second neural network which decodes the compressed vector z into a decompressed or reconstructed vector usually denoted ü, as close as possible to the input vector.
- the compressed vector z has a lower dimensionality than that of the input vector x and of the reconstructed vector ü: it is expressed by variables called latent variables, which are considered to define the essential characteristics of the vector.
- the decompressed vector ü is similar, but in general not strictly identical to the input vector x.
- Autoencoders typically provide very effective data dimensionality reduction, ignoring the "noise" of a signal.
- the compressed vector z and the latent variables can be considered to contain very high level information. For example, if the input vector x is an image, the elements of the compressed vector z can indicate different types of objects present in the image.
- variational autoencoders An evolution of autoencoders, called variational autoencoders, consists in representing each characteristic of the compressed vector z not as a single value, but as a probability distribution defined by a mean m and a standard deviation o. At decoding, a value is selected, for each characteristic, according to the probability distribution. Thus a generative model is obtained, since, for the same input vector x, a large number of decompressed vectors ü can be generated.
- Variational autoencoders are described for example by Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv: 1312.6114, or Diederik P. Kingma and Volodymyr Kuleshov. Stochastic Gradient Variational Bayes and the Variational Autoencoder. In ICLR, pp. 1-4, 2014.
- a new category of autoencoders is called multimodal autoencoders, multimodal autoencoders are described in particular by Wu, M., & Goodman, N. (2018). Multimodal generative models for scalable weakly-supervised learning.
- a multimodal autoencoder consists of a plurality of modalities, each modality comprising an encoder and a decoder. Each encoder encodes a separate input vector into a compressed vector, and then the compressed vectors provided by each of the modalities are merged into a single merged compressed vector for all modalities. Then, each decoder of each modality decodes the merged compressed vector into a decompressed vector specific to each modality.
- a multimodal autoencoder may or may not also be a variational autoencoder.
- the objective of multimodal autoencoders is to jointly train the encoders and decoders of the different modalities, in order to achieve extract latent variables summarizing the information provided by all the modalities.
- the different modalities can correspond to data provided by different sensors at the same time (for example, an RGB frame provided by camera in modality 1, a LIDAR point cloud in modality 2; or an RGB frame provided by a camera front of a vehicle in modality 1, an RGB frame supplied by a front camera of a vehicle in modality 2, etc.).
- the multimodal autoencoders thus make it possible, when they are applied to sensor data outputs, to perform a fusion of the sensor data making it possible to arrive at a common representation of the world by the various sensors.
- the latent variables of the merged compressed vector are thus particularly relevant for synthesizing the data provided by the sensors within the framework of a decision-making system, for example an autonomous vehicle.
- the combination by components consists in obtaining the values of each element of the merged compressed vector by combining the corresponding elements of the compressed vectors of each modality, that is to say that all the compressed vectors of the modalities, and merged have the same dimension, and the 1st element of the merged vector is obtained by combining all the 1st elements of the compressed vectors of the modalities, the 2nd element of the merged vector is obtained by combining all the 2 e elements of the compressed vectors of the modalities, etc.
- arXiv preprint arXiv:1911.03393., and the so-called “Robust Bayesian Committee Machines” method described by Deisenroth, M., & Ng, JW (2015, June). Distributed Gaussian processes. In International Conference on Machine Learning (pp. 1481-1490). PMLR.
- the subject of the invention is a multimodal latent data fusion autoencoder comprising: a plurality of modalities; a plurality of encoders encoding, for each modality of said plurality respectively, an input vector into a latent vector of predefined dimension, said predefined dimension being identical for each modality; a latent merging module merging the latent vectors encoded by said plurality of encoders into a merged vector of said predefined dimension; a plurality of decoders decoding, for each modality of said plurality of modalities respectively, said merged vector into an output vector, an output vector being of the same dimension as an input vector for each modality; said multimodal autoencoder being characterized in that the latent merging module is configured to assign to at least one element of the merged vector a value selected from among the element values of said latent vectors encoded by said plurality of encoders.
- the latent merging module is configured to assign to at least one element of the merged vector a value selected randomly from among the values of the elements of said latent vectors encoded by said plurality of encoders.
- the autoencoder is a variational autoencoder, and each element of one of said latent vectors encoded by said plurality of encoders, and of the merged vector is formed from a mean and a standard deviation.
- the latent fusion module is configured to assign to at least one element of the merged vector the mean and the standard deviation of the element having the smallest standard deviation among the corresponding elements of said latent vectors encoded by said plurality of encoders.
- each encoder of each modality of said plurality of modalities takes as input measurements of each sensor of a plurality of sensors respectively.
- the sensors of said plurality of sensors are co-located.
- the sensors of said plurality of sensors are located in a motor vehicle.
- the invention also relates to a method for training a multimodal latent data fusion autoencoder comprising a plurality of modalities, said method comprising several training iterations, each training iteration comprising: encoding , by a plurality of encoders, for each modality of said plurality respectively, of an input vector into a latent vector of predefined dimension, said predefined dimension being identical for each modality; latently merging the latent vectors encoded by said plurality of encoders into a merged vector of said predefined dimension; decoding, by a plurality of decoders, for each modality of said plurality of modalities respectively, said merged vector into an output vector, an output vector being of the same dimension as an input vector for each modality; the calculation of a loss function, from the output vectors and the input vectors; adaptation of encoders and decoders, to minimize said loss function; said training method being characterized in that the la
- the invention also relates to a method of using a multimodal latent data fusion autoencoder comprising a plurality of modalities, said method comprising: the encoding, by a plurality of encoders, for each modality of said plurality, respectively, of an input vector into a latent vector of predefined dimension, said predefined dimension being identical for each modality; the latent merging of the latent vectors encoded by said plurality of encoders into a merged vector of said predefined dimension; the decoding, by a plurality of decoders, for each modality of said plurality of modalities respectively, of said merged vector into an output vector, an output vector being of the same dimension as an input vector for each modality; said method of use being characterized in that the latent merging consists in assigning to at least one element of the merged vector a value selected from among the values of the elements of said latent vectors encoded by said plurality of encoders.
- the invention also relates to a computer program comprising program code instructions recorded on a computer-readable medium, said program code instructions being configured, when said program runs on a computer to execute a method according to one of the embodiments of the invention.
- the invention also relates to a computing system comprising: a plurality of sensors; at least one calculation unit configured to execute a multimodal latent data fusion autoencoder according to one of the embodiments of the invention, each sensor of said plurality being respectively associated with a modality of the plurality of modalities of the autoencoder .
- FIG.1a a first example of a calculation system according to a set of embodiments of the invention
- FIG.1b a second example of a calculation system according to a set of embodiments of the invention.
- FIG.2a a first example of a multimodal latent data fusion autoencoder according to a set of embodiments of the invention
- FIG.2b a second example of a multimodal variational autoencoder with latent data fusion according to a set of embodiments of the invention
- FIG.3 an example of a method for training a multimodal latent data fusion autoencoder according to one set of embodiments of the invention
- FIG. 4 an example of a method of using a multimodal latent data fusion autoencoder according to a set of embodiments of the invention
- FIG. 5a a first example of estimating the results of a latent merger, using a state-of-the-art method called “Product of Experts”;
- FIG. 5b a first example of estimating the results of a latent fusion, according to a set of embodiments of the invention
- FIG. 5c a second example of estimating the results of a latent fusion, according to a set of embodiments of the invention.
- FIG.6a a first example of multimodal reconstitution of data from camera, LIDAR and RADAR sensors embedded in a car;
- FIG.6b a second example of multimodal reconstruction of data from camera, LIDAR and RADAR sensors embedded in a car.
- FIG. 1a represents a first example of a calculation system according to a set of embodiments of the invention.
- the Sysa system is made up of an Auta motor vehicle.
- the Auta vehicle is equipped with a plurality of Capt1, Capt2, etc. sensors as well as at least one Calca calculation unit.
- a calculation unit can be a processor operating according to software instructions, a hardware configuration of a processor or a combination thereof. It should be noted that any or all of the functions described herein may be implemented in a pure hardware implementation and/or by a processor operating in accordance with software instructions. It should also be understood that any or all of the software instructions may be stored on a non-transitory computer readable medium. For the sake of simplicity, in the rest of the description, the calculation unit or units may be referred to as the “calculation unit”. However, it is at note that the operations of the invention can also be performed in a single calculation unit, or a plurality of calculation units, for example a plurality of processors or a multi-core processor.
- the Calca calculation unit is configured to run an AE latent data fusion multimodal autoencoder comprising a plurality of modalities.
- Each modality of the multimodal encoder is respectively associated with a sensor among the plurality of sensors Capt1, Capt2, etc. In other words, the outputs of each of the sensors are provided as input to the different modes of the AE autoencoder.
- one of the objectives of the invention, and of the multimodal autoencoder AE is to obtain a latent representation of the autoencoder which provides a representation low dimensionality of the data captured by all the sensors.
- different combinations of sensors can be used for said plurality, such as for example:
- the AE autoencoder can take as input data provided by all the sensors present in the Auta vehicle, or only from some of them.
- the Sysa system can be used both for the training phases and the inference phases of the AE autoencoder.
- the at least one calculation unit Calca is further configured to execute an autonomous guidance module of the Auta vehicle.
- the use of the autoencoder AE advantageously makes it possible to perform a very efficient fusion of data from the sensors, and to result in a representation of low dimensionality of all the data captured by the sensors. This representation of low dimensionality allows thus to obtain high-level characteristics on the environment of the vehicle, and to facilitate the autonomous guidance of the latter.
- FIG. 1b represents a second example of a calculation system according to a set of embodiments of the invention.
- the Sysb system is formed by a motor vehicle Autb, and a computing device Dispb communicating via a Transb data transmission link.
- the calculation device Dispb can be any device capable of performing calculations, such as a server, a personal computer, a tablet or a smartphone.
- the calculation device Dispb comprises at least one calculation unit Calcb capable of executing a multimodal encoder AE.
- the data transmission link Transb can be formed of any element allowing a data link between the motor vehicle Autb and the computing device Dispb. For example, a 4G, Wi-Fi or Bluetooth type data link can be used.
- the data transmission link Transb notably enables the vehicle Autb to transmit the measurements of the sensors to the calculation device Calcb for the processing thereof.
- the motor vehicle Autb comprises the plurality of sensors Capt1, Capt2, etc., respectively associated with the modalities of the autoencoder AE.
- Sysa and Sysb systems therefore resides in the fact that, in the Sysa system, the execution of the autoencoder takes place in the Auta vehicle itself, whereas, within the framework of the system Sysb, it is carried out remotely in a remote device Dispb.
- Sysa and Sysb systems are provided by way of example only.
- a multimodal autoencoder according to the invention is able to perform a fusion of data from several sources to arrive at a low-dimensional representation of the data representative of the all sources, whether the data sources are sensors or not. More particularly, an autoencoder according to the invention is able to perform a fusion of sensor measurements. The invention is therefore applicable to a plurality of sensors, whether or not they are located in a motor vehicle.
- the invention is in particular applicable to collocated sensors, that is to say located substantially at the same location, since such a plurality of sensors will measure data in the same environment, on which an effective data fusion can be performed.
- Collocated sensors can for example be sensors located in the same box, the same vehicle, or more generally sensors located close to each other, whether or not they are located in a motor vehicle.
- Figure 2a shows a first example of a multimodal latent data fusion autoencoder according to one set of embodiments of the invention.
- the multimodal latent data fusion autoencoder AEa comprises a plurality of modalities Mod1, Mod2, Mod3. Although three modalities are represented in Figure 2a, this example is not limiting, and any number of modalities equal to or greater than 2 can be used.
- Each modality comprises an encoder Enc1 a, Enc2a, Enc3a, encoding an input vector x lt x 2 , x 3 into a latent vector z lt z 2 , z 3 , also called compressed vector of predefined dimension n, said predefined dimension being identical for each modality.
- the 1 st modality Mod1 comprises an encoder Enc1 a encoding an input vector x t into a latent vector z x
- the 2nd modality Mod2 comprises an encoder Enc2a encoding an input vector x 2 into a latent vector z 2 ;
- the 3rd modality Mod3 comprises an encoder Enc3a encoding an input vector x 3 into a latent vector z 3 ;
- the latent vectors z lt z 2 , z 3 have the same predefined dimension n.
- the input vectors x lt x 2 , x 3 correspond to different data sources.
- the input vector x t can correspond to the measurements from the Capt1 sensor, and the input vector x 2 to the measurements from the Capt2 sensor.
- the input vectors x lt x 2 , x 3 can correspond to raw data coming from the sensors and/or to pre-processed data.
- an input vector can correspond to an image, or an image on which a semantic segmentation has been performed.
- the AEa multimodal autoencoder comprises a Fusa latent fusion module merging the latent vectors output from said encoders into a merged vector z of said predefined dimension.
- the Fusa latent fusion module takes as input latent vectors z lt z 2 , z 3 generated by the encoders for each modality, and merges them into a single merged vector z.
- the merged vector z therefore provides a high level, low dimensionality representation of the information contained in the set of input vectors x lt x 2 , x 3.
- the multimodal data fusion autoencoder AEa further comprises, for each modality of said plurality, a decoder Declb, Dec2b, Dec3b decoding the merged vector z into an output vector x ⁇ , x 2 , x 3 likewise dimension than the input vector of said modality.
- Mod1 comprises a Decla decoder decoding the merged vector z into an output vector x t of the same dimension as the input vector x t ;
- the 2nd modality Mod2 comprises a decoder Dec2a decoding the merged vector z into an output vector x 2 of the same dimension as the input vector x 2 ;
- the 3rd modality Mod3 comprises a decoder Dec3a decoding the merged vector z into an output vector x 3 of the same dimension as the input vector x 3.
- the output vectors x x , x 2 , x 3 would respectively be similar to the input vectors x lt x 2 , x 3
- the merged vector z is enriched with information coming from the set of input vectors.
- the values of each of the output vectors x ⁇ , x 2 , x 3 therefore depend on the set of input vectors x lt x 2 , x 3 .
- Each of the encoders Enc1 a, Enc2a, Enc3a, and the decoders Decl b, Dec2b, Dec3b is a neural network.
- the characteristics of each of these neural networks depend on the topology of the input data.
- convolutional neural networks can be used, in particular for image data from cameras. Those skilled in the art will be able to effortlessly determine the most suitable type of neural network for a given type of input.
- the following types of neural networks can be used:
- FCN Fully-Connected Networks
- recursive networks for example, a recurrent network with short and long term memory (in English Long Short Term Memory or LSTM) and Recurrent Neural Networks (in English Recurrent Neural Networks or RNN);
- the Fusa latent fusion module is configured to assign to at least one element of the merged vector a value selected from among the values of the corresponding elements of the latent vectors of said modalities.
- each element of index i (with ie [l; n] if the indices of the elements are denoted 1 to n) of the merged vector z is chosen as one of the values of the elements of index i latent vectors z lt z 2 , z 3 .
- the vector selected for the value is obviously not necessarily the same for each element.
- the value of the 1st element of z can be selected like that of the 1st element of z 2 , but that of the 2nd element of z can be selected like that of the 2nd element of z lt etc.
- the choice of the latent vector from which the value of a given element of the merged vector is selected can be different each time the multimodal autoencoder is used. In particular, the selection can be re-performed for each element, independently, at each iteration of a training (or learning) phase of the autoencoder.
- the invention makes it possible to achieve a very effective fusion of the input data, and the learning of the different modes of the autoencoder AEa.
- the selection of the value to be assigned to each element of the merged vector z from among the values of the corresponding elements of z lt z 2 , z 3 is carried out randomly.
- the value of the 1 st element of z is chosen randomly from among the values of the 1 st elements of z lt z 2 , z 3 ;
- the value of the 2 nd element of z is chosen randomly from among the values of the 2 nd elements of z lt z 2 , z 3 ;
- This method can be called “Indian roulette”. This makes it possible, in a particularly effective manner, to cause the encoders and decoders of the different modalities to code, for each element of their compressed vectors, the same latent variables. Indeed, each training iteration will take into account, randomly, a value from one of the encoders for each element of z, which makes it possible to train the encoders and decoders so that the same characteristic is coded for a given element, regardless of the mode selected.
- FIG. 2b represents a second example of a multimodal variational autoencoder with latent data fusion according to a set of embodiments of the invention.
- the multimodal autoencoder AEb comprises a plurality of modalities, denoted in this example Mod1, Mod2, Mod3, each modality taking as input an input vector x lt x 2 , x 3 encoded respectively by an encoder Enc1 b, Enc2b, Enc3b into a compressed vector, or latent z lt z 2 , z 3 , then the fusion module Fusb merges the latent vectors z lt z 2 , z 3 into a merged vector z which will be decoded for each modality by a decoder Decl b, Dec2b, Dec3b respectively into an output vector x ⁇ , x 2 , x 3 .
- the autoencoder AEb is a variational autoencoder, that is to say that each element of the latent vectors z lt z 2 , z 3 , and of the merged vector z comprises, not not a single value, but a mean m and a standard deviation o defining a distribution.
- a value is selected from among the probability distribution defined by the values of mean m and standard deviation s 2 for each element of z, thus allowing decoding.
- the latent fusion module Fusb is configured to assign to at least one element of the merged vector the mean and the standard deviation of the element having the deviation- weakest type among the corresponding elements of the latent vectors of said modalities.
- This method can be called “survival of the fittest”. This makes it possible to select, for each latent variable, the corresponding value having the lowest standard deviation among the latent vectors z lt z 2 , z 3 , and therefore to have the most precise values for each latent variable .
- FIG. 3 represents an example of a training method for a multimodal latent data fusion autoencoder according to a set of embodiments of the invention.
- the method 300 is a method for training a multimodal autoencoder with latent data fusion such as the autoencoders AE, AEa, AEb.
- the method includes a plurality of iterations of the following steps.
- Each iteration comprises a first step 310 of encoding, by a plurality of encoders such as the encoders Enc1 a, Enc1 b, Enc2a, Enc2b, Enc3a, Enc3b, for each modality of said plurality respectively, of a vector input such as the vector x lt x 2 or x 3 into a latent vector such as the vector z lt z 2 or z 3 of predefined dimension n, said predefined dimension being identical for each modality.
- Each iteration comprises a second step 320 of latent merging of the latent vectors encoded by said plurality of encoders into a merged vector z of said predefined dimension.
- Each iteration comprises a third step 330 of decoding, by a plurality of decoders such as the Decla, Declb, Dec2a, Dec2b, Dec3a, Dec3b decoders, for each modality of said plurality of modalities respectively, of said vector merged into an output vector such as the vector x lt x 2 or x 3 , an output vector being of the same dimension as an input vector for each modality.
- a plurality of decoders such as the Decla, Declb, Dec2a, Dec2b, Dec3a, Dec3b decoders
- Each iteration includes a fourth step 340 of calculating a loss function, from the output vectors and the input vectors.
- the loss function makes it possible to evaluate the loss between the input vectors x lt x 2 , x 3 , and the output vectors x ⁇ , x 2 , x 3.
- different loss functions can be used such as Euclidean distances (for example L2 or L1 type functions) or functions linked to the perception of images (in English pixel perceptual loss functions).
- Each iteration includes a fifth step 350 of adaptation of the encoders and decoders, to minimize the loss function.
- This fifth step can in particular be carried out by the so-called gradient backpropagation method, to adapt the weights and biases of the neural networks forming the encoders and decoders, from the gradient of the loss function.
- the training method is characterized in that the latent merging consists in assigning to at least one element of the merged vector a value selected from among the values of the elements of said latent vectors encoded by said plurality of encoders.
- steps 310, 320, 330, 340, and 350 are performed in order to perform the training.
- a predefined number of iterations can for example be performed, the iterations can continue until the loss function is below a threshold, or until the difference between the values of the loss function loss between two successive iterations is less than a threshold.
- FIG. 4 represents an example of a method of using a multimodal autoencoder with latent data fusion according to a set of embodiments of the invention.
- the method of use 400 corresponds to the actual use of a multimodal latent data fusion autoencoder such as one of the autoencoders AE, AEa, AEb, i.e. it corresponds to an inference phase.
- the method 400 comprises a first step 410 of encoding, by a plurality of encoders such as the encoders Enc1a, Enc1b, Enc2a, Enc2b, Enc3a, Enc3b, for each modality of said plurality respectively, of a input vector such as the vector x lt x 2 , x 3 into a latent vector such as the vector z lt z 2 , z 3 of predefined dimension n, said predefined dimension being identical for each modality.
- a plurality of encoders such as the encoders Enc1a, Enc1b, Enc2a, Enc2b, Enc3a, Enc3b, for each modality of said plurality respectively, of a input vector such as the vector x lt x 2 , x 3 into a latent vector such as the vector z lt z 2 , z 3 of predefined dimension n, said predefined dimension being identical for each modality.
- the method 400 comprises a second step 420 of latent merging of the latent vectors encoded by said plurality of encoders into a merged vector z of said predefined dimension.
- the method 400 comprises a third step 430 of decoding, by a plurality of decoders such as the decoders Decla, Decl b, Dec2a, Dec2b, Dec3a, Dec3b, for each modality of said plurality of modalities respectively, of said vector merged into an output vector such that the vectors x ⁇ , x 2 , x 3 , an output vector being of the same dimension as an input vector for each modality.
- a plurality of decoders such as the decoders Decla, Decl b, Dec2a, Dec2b, Dec3a, Dec3b
- the method 400 is characterized in that the latent merging consists in assigning to at least one element of the merged vector a value selected from among the values of the elements of the latent vectors encoded by said plurality of encoders.
- FIG. 5a represents a first example of estimating the results of a latent fusion, according to a state-of-the-art method called “Product of Experts”;
- FIG. 5b a first example of estimating the results of a latent fusion, according to a set of embodiments of the invention
- FIG. 5c a second example of estimating the results of a latent fusion, according to a set of embodiments of the invention
- the graphs 500a, 500b, 500c represent three examples of estimating the results of a latent fusion, respectively for:
- the autoencoders were trained in the same way (same training set, number of iterations and loss function for training in particular), only the latent fusion being modified between the three examples.
- the graphs 500a, 500b, 500c are based on the same example, in which the position of a toy in an image is estimated from more or less noisy images.
- the horizontal axis represents the noise level (or SNR, from English Signal to Noise Ratio, or in French Radio Signal to Noise), in dB of the input image, and the vertical axis the variance of the results, with a logarithmic scale. Better fusion quality is associated with lower variance (thus lower on the vertical axis).
- FIG. 6a represents a first example of multimodal reconstruction of data from camera, LIDAR and RADAR sensors on board a car.
- FIG. 6b represents a second example of multimodal reconstruction of data from camera, LIDAR and RADAR sensors on board a car.
- a scene is captured by three sensors (respectively camera, LIDAR and RADAR) embedded in the same car.
- a first scene consists of 3 objects in front of two walls forming an angle.
- a second scene contains a single object in front of the same two walls.
- the thumbnails 610a, 610b represent the object or objects seen by the LIDAR in front of the walls.
- the car where the LIDAR is located is located at the bottom right of this representation, as indicated by the apparent shooting cone at the bottom of the thumbnails.
- the thumbnails 611 a, 611 b represent an image of the scene seen by a camera on board the same of the car.
- Thumbnails 612a, 612b represent a RADAR view of the same scene.
- the view is a 2D view of the "top view” location and shape items.
- the view is also reversed with respect to the other thumbnails, ie the “L” at the bottom of these thumbnails represents the two walls, and the line(s) the object(s).
- a fusion of multimodal data is performed by a multimodal autoencoder according to one of the embodiments of the invention.
- the vectors representing the shots 610a, 611a, 612a, 610b, 611b, 612b are therefore encoded into a latent vector for each of the two scenes, then the latent vectors are decoded into decompressed vectors.
- each latent vector has 64 elements.
- the thumbnails 620a, 620b, 621a, 621b, 622a, 622b respectively represent the LIDAR, camera and RADAR views of the first and the second scene as restored by the decompressed vectors. These decompressed views are very close to the initial views, even though they were generated from a latent vector comprising a very small amount of information (64 elements in this example).
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Image Processing (AREA)
- Compression, Expansion, Code Conversion, And Decoders (AREA)
Abstract
Description
Claims
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| FR2103265A FR3121535B1 (fr) | 2021-03-30 | 2021-03-30 | Autoencodeur multimodal a fusion de donnees latente amelioree |
| PCT/EP2022/058163 WO2022207573A1 (fr) | 2021-03-30 | 2022-03-28 | Autoencodeur multimodal a fusion de donnees latente amelioree |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| EP4315170A1 true EP4315170A1 (fr) | 2024-02-07 |
Family
ID=76601315
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP22719247.3A Pending EP4315170A1 (fr) | 2021-03-30 | 2022-03-28 | Autoencodeur multimodal a fusion de donnees latente amelioree |
Country Status (3)
| Country | Link |
|---|---|
| EP (1) | EP4315170A1 (fr) |
| FR (1) | FR3121535B1 (fr) |
| WO (1) | WO2022207573A1 (fr) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CA3096285A1 (fr) | 2020-10-16 | 2022-04-16 | Pulsemedica Corp. | Dispositif, systeme et methodes d'administration de laser et d'imagerie ophtalmologique |
| CA3100460A1 (fr) | 2020-11-24 | 2022-05-24 | Pulsemedica Corp. | Ciblage de modulation d'eclairage spatial de lasers therapeutiques pour le traitement de conditions ophtalmologiques |
| US12032111B2 (en) | 2021-03-05 | 2024-07-09 | Saudi Arabian Oil Company | Method and system for faster seismic imaging using machine learning |
| US12013508B2 (en) * | 2021-10-28 | 2024-06-18 | Saudi Arabian Oil Company | Method and system for determining seismic processing parameters using machine learning |
| WO2024246714A1 (fr) * | 2023-05-26 | 2024-12-05 | Pulsemedica Corp. | Systèmes, procédés et dispositifs de fusion multimodale |
| CN116843963A (zh) * | 2023-06-28 | 2023-10-03 | 北京百度网讯科技有限公司 | 图像识别方法、装置、设备及存储介质 |
| CN119026004A (zh) * | 2024-08-05 | 2024-11-26 | 中国电子科技集团公司第十研究所 | 一种跨域多模态数据自动关联映射与融合识别方法 |
| US20260094101A1 (en) * | 2024-09-30 | 2026-04-02 | Intuit Inc. | Machine learning framework that extracts actionable insights from disparate data sources |
| CN120597220B (zh) * | 2025-08-08 | 2025-12-16 | 中国电子科技集团公司第五十四研究所 | 一种光电-雷达时空协同的脉冲路由识别方法 |
-
2021
- 2021-03-30 FR FR2103265A patent/FR3121535B1/fr active Active
-
2022
- 2022-03-28 WO PCT/EP2022/058163 patent/WO2022207573A1/fr not_active Ceased
- 2022-03-28 EP EP22719247.3A patent/EP4315170A1/fr active Pending
Also Published As
| Publication number | Publication date |
|---|---|
| FR3121535B1 (fr) | 2024-05-24 |
| WO2022207573A1 (fr) | 2022-10-06 |
| FR3121535A1 (fr) | 2022-10-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| EP4315170A1 (fr) | Autoencodeur multimodal a fusion de donnees latente amelioree | |
| EP3301617B1 (fr) | Procédés d'apprentissage sécurisé de paramètres d'un réseau de neurones à convolution, et de classification sécurisée d'une donnée d'entrée | |
| US10929676B2 (en) | Video recognition using multiple modalities | |
| EP4256412B1 (fr) | Système et procédé de contrôle de véhicule à base d'apprentissage machine | |
| FR3112007A1 (fr) | Procédé et appareil d’entraînement de modèle, et procédé et appareil de prédiction | |
| EP3221841B1 (fr) | Procede et dispositif de filtrage adaptatif temps reel d'images de disparite ou de profondeur bruitees | |
| CN116547696A (zh) | 一种图像增强方法和装置 | |
| US20230410255A1 (en) | Decreased quantization latency | |
| WO2022008605A1 (fr) | Dispositif électronique et procédé de traitement de données à base de réseaux génératifs inversibles, système électronique de détection et programme d'ordinateur associés | |
| Bach et al. | Underwater image enhancement with physical-based denoising diffusion implicit models | |
| EP4394658A1 (fr) | Procédé amélioré d apprentissage sensible à la quantification pour un réseau de neurones | |
| EP4187445A1 (fr) | Procédé d'apprentissage de valeurs de poids synaptique d'un réseau de neurones, procédé de traitement de données, programme d'ordinateur, calculateur et système de traitement associés | |
| US20250217988A1 (en) | Image segmentation mask refinement with diffusion model | |
| WO2020157733A1 (fr) | Procédé dynamique d'imagerie tridimensionnelle | |
| EP2943935B1 (fr) | Estimation de mouvement d'une image | |
| EP4191530A1 (fr) | Procédé de localisation et cartographie simultanées intégrant un masquage temporel auto-supervisé et modèle d'apprentissage automatique pour générer un tel masquage | |
| EP2804129A1 (fr) | Procédé de reconnaissance vocale visuelle avec sélection de groupes de points d'intérêts les plus pertinents | |
| EP3797509B1 (fr) | Traitement d'un bruit impulsionnel dans une séquence vidéo | |
| EP4202770A1 (fr) | Reseau de neurones avec generation a la volee des parametres du reseau | |
| EP4012619A1 (fr) | Méthode de compression d'un réseau de neurones artificiel | |
| FR3120173A1 (fr) | Détermination d’au moins un mode de codage d’image ou d’au moins un mode de décodage d’image, codage et décodage d’image utilisant une telle détermination | |
| Thakur et al. | USRNet: A simple yet effective Underwater Scene Restoration Network | |
| FR3162093A1 (fr) | Procédé et dispositif de généralisation d’un modèle de classification d’image pour système d’aide à la conduite d’un véhicule | |
| Fabbri | Enhancing Visual Perception in Noisy Environments using Generative Adversarial Networks | |
| FR3160033A1 (fr) | Procédé et dispositif de segmentation d’objets matériels dans un espace environnant un robot autonome |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: UNKNOWN |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE |
|
| PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE |
|
| 17P | Request for examination filed |
Effective date: 20230915 |
|
| AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR |
|
| RAP1 | Party data changed (applicant data changed or rights of an application transferred) |
Owner name: AMPERE SAS |
|
| DAV | Request for validation of the european patent (deleted) | ||
| DAX | Request for extension of the european patent (deleted) | ||
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: EXAMINATION IS IN PROGRESS |
|
| 17Q | First examination report despatched |
Effective date: 20250731 |