Abstract: A block of video data is split using one or more of several possible partition operations by using the partitioning choices obtained through use of a texture-based image partitioning. In at least one embodiment, the block is split in one or more splitting operations using a convolutional neural network. In another embodiment, inputs to the convolutional neural network come from pixels along the block's causal borders. In another embodiment, boundary information, such as the location of partitions in spatially neighboring blocks, is used by the texture analysis. Methods, apparatus, and signal embodiments are provided for encoding.

Abstract: Encoders and decoders of digital video signals use combined intra prediction modes for some images. In at least one embodiment, combined intra prediction modes can result from the intra prediction modes of neighboring blocks. The combined intra prediction mode can be added to a most probable modes list. Various embodiments describe techniques for combining the intra prediction modes, comprising a combination of reference samples and a combination of directions to form a prediction. Another embodiment adds a combination mode to the most probable modes list if the two modes that are combined to form it are directional modes with angle difference less than 90 degrees. Another embodiment uses a linear combination of intra prediction modes and another embodiment performs a linear combination which depends on the distance of a prediction from the left and above blocs.

Abstract: The present disclosure relates to a method including encoding a data set in a signal, the encoding comprising quantizing the data set by using a codebook obtained by clustering the data set, the clustering taking account of a probability of appearance of data in the dataset; the probability being bounded to a bounding value. The present disclosure also relates to a method including encoding in a signal a first weight of a layer of a Deep Neural Network, the encoding taking into account an impact of a modification of a second weight on an accuracy of the Deep Neural Network.

Abstract: The method for intra frame prediction allows for the recovery of some intra directional modes that are otherwise rendered useless due to the classical padding process. Directional modes are recovered by filling pixel locations in the prediction unit with samples from the portion of the unit that is determined to have samples available for use using a counterpart directional mode to the one that is useless. When it is determined that reference samples are missing for a particular prediction unit or portion thereof given a particular directional mode, a determination is made about whether reference samples along another portion of the prediction unit are available. The other portion is the side of the prediction unit at an opposite end of the particular directional mode.

Abstract: Decoding or encoding picture information can involve determining a value associated with a prediction of a pixel of a block of picture information, wherein, the prediction is based on an intra-prediction along a non-diagonal direction, determining, from among a plurality of decoded pixels neighboring the block on the top and the left, at least four reference samples for the block based on the non-diagonal direction, and wherein the determining of the value comprises an interpolation based on a quadratic model and the four reference samples; and decoding or encoding at least a portion of the picture information based on the value.

Abstract: The present disclosure relates to a method including reshaping a first tensor of weights, by using one or more second tensor having a lower dimension than the first tensor dimension and encoding the second tensor in a signal The present disclosure relates to a method including obtaining a first tensor of weights by reshaping one or more second tensor hav ing a lower dimension than the first tensor dimension, the one or more second tensor being decoded from a signal. The present disclosure further relates to the corresponding dev ices, signal, and computer readable storage media.

Abstract: A method for encoding a block is disclosed. To this aim, a split mode is determined based on a rate-distortion optimization using a texture-based split prediction set obtained for the block. As an example, the split mode is determined by adapting the texture-based split prediction set according to at least one of a binary or triple split mode non-redundancy constraint or a heuristic-based split mode set pruning. The block is finally encoded using the determined split mode.

Abstract: Methods and apparatuses for video coding and decoding are provided. The method of video encoding includes accessing a bin of a syntax element associated with a block in a picture of a video, determining a context for the bin of the syntax element associated with the syntax element and entropy encoding the bin of the syntax element based on the determined context wherein either the bin of the syntax element is based on the relevance of a prediction by a neural network of the syntax element or the probability associated to the context is determined by a neural network. A bitstream formatted to include encoded data, a computer-readable storage medium and a computer-readable program product are also described.

Abstract: Different implementations are described, particularly implementations for video encoding and decoding based on linear weighted intra prediction, also called matrix based intra prediction, are presented. Accordingly, the encoding or decoding comprises obtaining intra predicted samples of a block from a selected weight matrix and associated bias and from a set of neighboring reference samples; wherein any coefficient of the selected weight matrix is a power of two. According to relaxed version, not all but only a part of the coefficients of the selected weight matrix is a power of two. Advantageously, such arrangement allows to reduce the amount of memory for storing data and to reduce the complexity of the intra prediction samples computation.

Abstract: Different implementations are described, particularly implementations for video encoding and decoding based on linear weighted intra prediction, also called matrix based intra prediction, are presented. Accordingly, for a block being encoded or decoded in linear weighted intra prediction, obtaining intra predicted samples from at least two matrix-vector products between at least two selected weight matrices of reduced size and a set of neighboring reference samples. Advantageously, such arrangement allows to reduce the amount of memory for storing data and to reduce the complexity of the intra prediction samples computation.

Abstract: A block of video data is split using one or more of several possible partition operations by using the partitioning choices obtained through use of a texture-based image partitioning. In at least one embodiment, the block is split in one or more splitting operations using a convolutional neural network. In another embodiment, inputs to the convolutional neural network come from pixels along the block's causal borders. In another embodiment, boundary information, such as the location of partitions in spatially neighboring blocks, is used by the texture analysis. Methods, apparatus, and signal embodiments are provided for encoding.

Abstract: Different implementations are described for determining one or more illumination compensation parameters for a current block being encoded by a video encoder or decoded by a video decoder, based on the selection of one or more neighboring samples. The selection of the one or more neighboring samples is based on information used to reconstruct a plurality of neighboring reconstructed blocks. The selection may be based on the motion information, such as motion vector and reference picture information. In one example, only samples from neighboring reconstructed blocks that have (1) the same reference picture index and/or (2) a motion vector close to the motion vector of the current block is selected. In another example, if the current block derives or inherits some motion information from a top or left neighboring block, then only the top or left neighboring samples are selected for IC parameter calculation.

Abstract: At least one of the present embodiments generally relates to a method and an apparatus for video encoding and decoding, and more particularly, to a method and an apparatus for low-complexity bi-directional intra prediction. For example, a subset of a plurality of directional intra prediction modes is selected based on a shape of a current block of a picture being encoded or decoded. For a selected intra prediction mode in this subset, a sample value is bi-directionally intra predicted by using two predictors. The two predictors are substantially on a line based on a direction corresponding to the selected directional intra prediction mode. The bi-directional intra prediction may be only applied to large luma blocks, but not chroma blocks in order to reduce the computational complexity. When bi-directional intra prediction is used, the smoothing filtering may be omitted in intra prediction to further reduce the complexity.

Abstract: A method and apparatus for conveying information in a bitstream for deep neural network compression, such as in matrices representing weights, biases and non-linearities, to iteratively compress a pre-trained deep neural network by low displacement rank based approximation of the network layer weight matrices. The low displacement rank approximation allows for replacement of an original layer weight matrices of the pre-trained deep neural network as the sum of small number of structured matrices, allowing compression and low inference complexity. A decoder stage parses a bitstream for inference.

Abstract: For a picture with two or more color components, the prediction residuals for the first color component of a block to be encoded may be transformed with a first transform. The transform coefficients for the first color component may go through quantization, de-quantization and inverse transform to obtain reconstructed prediction residuals. Based on the reconstructed prediction residuals for the first color component, the phases of the transform basis function of the first transform can be adjusted to improve the sparsity of the transformed signal. The prediction residuals for the remaining color components may then be transformed with the adjusted transform. In order to determine the phase shift factor, the reconstructed prediction residuals for the first color component may be transformed with the first transform, adjusted by different candidate phase shift factors, and the candidate phase shift factor that provides a smallest sparsity measure can be selected for the block.

Abstract: A block of video data is split using at least one splitting rule and coded using existing transform sizes through one of several embodiments. In one embodiment, the block is split using a triple tree splitting operation if said at least two rectangular sub-blocks can be achieved with two successions of split operations with at least one triple tree split. In another embodiment, the video block is split using an asymmetric binary tree splitting if said at least two rectangular sub-blocks can be achieved with two successions of split operations with at least one asymmetric binary split. In another embodiment, a video block is split using successive splits, using both of the rules of the other embodiments.

Abstract: When a block is intra predicted, the predicted samples on the boundary are subject to boundary filtering for certain modes. Various embodiments are proposed to perform boundary filtering for intra prediction in the DC or PLANAR mode, using a gradient based approach. The gradient for a particular reference sample is calculated based on a difference between the particular reference sample and a predictor for predicting the particular reference sample. For the PLANAR mode, the difference between a top (left) reference sample and a linear interpolation using the top (left) reference samples is used as the gradient, a fractional part of which is propagated to the initial predicted samples with a decay function. For the DC mode, the difference between a top (left) reference sample and an average of reference samples is used as the gradient. The average can be computed with the top reference samples and/or left reference samples.

Abstract: A method and apparatus for performing prediction for encoding or decoding uses intra prediction with sub-partitions. The sub-partitions are oriented either horizontally or vertically and can use wide-angle modes different than that of the video block they originate from. Reference samples for a sub-partition are those of the video block when reference samples for the sub-partition are not available, such as due to direction. In an embodiment, when a sub-partition is square, conventional intra prediction directions are used. Reference samples can be used from a block above and right of the video block or left and below the video block using a mapping along a prediction direction, vertically, or horizontally.

Abstract: A method and an apparatus for performing deep neural network compression use an approximation training set along with information, such as in matrices representing weights, biases and non-linearities, to iteratively compress a pre-trained deep neural network by low displacement rank based approximation of the network layer weight matrices. The low displacement rank approximation allows for replacement of an original layer weight matrices of the pre-trained deep neural network as the sum of a small number of structured matrices, allowing compression and low inference complexity.

Abstract: The method for intra frame prediction allows for the recovery of some intra directional modes that are otherwise rendered useless due to the classical padding process. Directional modes are recovered by filling pixel locations in the prediction unit with samples from the portion of the unit that is determined to have samples available for use using a counterpart directional mode to the one that is useless. When it is determined that reference samples are missing for a particular prediction unit or portion thereof given a particular directional mode, a determination is made about whether reference samples along another portion of the prediction unit are available. The other portion is the side of the prediction unit at an opposite end of the particular directional mode.