Motion flow coding for deep learning based YUV video compression
Video encoding and decoding is implemented with auto encoders using luminance information to derive motion information for chrominance prediction. In one embodiment YUV 4:2:0 video is encoded and decoded in which luminance information is downsampled to generate predictions from chrominance components of a reference frame. In a related embodiment, more than one reference frame is used for predictions. In another embodiment, convolutions and transpose convolutions implement derivation of motion information.
1 . A method, comprising: concatenating luminance data of a current and at least one reference block; determining motion information from the concatenated luminance data; processing said determined motion information with a tanh function and a warping to generate a motion tensor; encoding said determined motion information; down sampling said encoded motion information; determining a prediction of chrominance data based on said downsampled encoded motion information, said motion tensor and chrominance data of the at least one reference block; determining a prediction of luminance data based on said encoded motion information, said motion tensor and luminance data of the at least one reference block; determining residual data of the current block based on said predictions; and, encoding said residual data into a bitstream also comprising said encoded motion information.
2 . An apparatus, comprising: a processor, configured to: concatenate luminance data of a current and at least one reference block; determine motion information from the concatenated luminance data; process said determined motion information with a tanh function and a warping to generate a motion tensor; encode said determined motion information; down sample said encoded motion information; determine a prediction of chrominance data based on said downsampled encoded motion information, said motion tensor and chrominance data of the at least one reference block; determine a prediction of luminance data based on said encoded motion information, said motion tensor and luminance data of the at least one reference block; determine residual data of the current block based on said predictions; and, encode said residual data into a bitstream also comprising said encoded motion information.
3 . A method, comprising:
decoding motion information from a bitstream;
down sampling said decoded motion information;
processing said decoded motion information with a tanh function and a warping to generate a motion tensor;
determining a prediction of chrominance data based on said downsampled decoded motion information, said motion tensor and chrominance data of at least one reference block;
determining a prediction of luminance data based on said decoded motion information, said motion tensor and luminance data of the at least one reference block;
decoding residual information; and,
reconstructing a video block using said decoded residual information and said predictions.
4 . An apparatus, comprising:
a processor, configured to:
decode motion information from a bitstream;
down sample said decoded motion information;
process said decoded motion information with a tanh function and a warping to generate a motion tensor;
determine a prediction of chrominance data based on said downsampled decoded motion information, said motion tensor and chrominance data of at least one reference block;
determine a prediction of luminance data based on said decoded motion information, said motion tensor and luminance data of the at least one reference block;
decode residual information; and,
reconstruct a video block using said decoded residual information and said predictions.
5 . The method of claim 1 , wherein said encoding is performed using an auto encoder.
6 . The method of claim 1 , wherein said prediction comprises:
computing a Gaussian volume on at least one reference frame;
performing a tanh transformation on motion information; and,
performing a warp function on said tanh function output to generate a prediction.
7 . The method of claim 6 wherein said warping is separate for luminance and chrominance to generate a prediction.
8 . The method of claim 1 , wherein said luminance and chrominance data is in YUV 4:4:4 format.
9 . The apparatus of claim 2 , wherein a parametric model is used on the at least one reference block to generate a new block.
10 . A device comprising:
an apparatus according to claim 1 ; and
at least one of (i) an antenna configured to receive a signal, the signal including the video block, (ii) a band limiter configured to limit the received signal to a band of frequencies that includes the video block, and (iii) a display configured to display an output representative of a video block.
11 . A non-transitory computer readable medium containing data content generated according to the method of claim 1 , for playback using a processor.
12 . A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of claim 1 .
13 . The method of claim 3 , wherein said encoding and decoding is performed using an auto encoder.
14 . The method of claim 3 , wherein said prediction comprises:
computing a Gaussian volume on at least one reference frame;
performing a tanh transformation on motion information; and,
performing a warp function on said tanh function output to generate a prediction.
15 . The method of claim 3 , wherein said decoding of motion information comprises at least one transpose convolution and at least one rectified linear unit function.
16 . The method of claim 14 wherein said warping is separate for luminance and chrominance to generate a prediction.
17 . The apparatus of claim 4 , wherein said luminance and chrominance data is in YUV 4:4:4 format.
18 . The apparatus of claim 4 , wherein a parametric model is used on the at least one reference block to generate a new block.
19 . The method of claim 15 , wherein said decoding is split into multiple branches for at least one convolutional layer.