Systems and methods for compression of feature data using joint coding in coding of multi-dimensional data

1 . A method of compressing feature data corresponding to video data, the method comprising:

for each of N pictures included in video data, generating feature data including a number of channels, C, corresponding to a scale, such that the generated feature data for each of the N pictures includes a feature tensor including a channel dimension, a height dimension, and a width dimension, (C×H×W);

concatenating the generated feature data about the channel dimension, such that the concatenated feature data includes a concatenated feature tensor including a channel dimension, wherein the number of channels in the concatenated feature tensor is given by N multiplied by the number of channels corresponding to the scale (N×C);

reducing the number of channels in the concatenated feature tensor to generate reduced concatenated feature data, wherein reducing the number of channels in the concatenated feature tensor includes:

performing a sequence of res2d operations on the concatenated feature tensor to generate a refinement value, wherein each of the res2d operations is a residual block of a compressive autoencoder,

adding the refinement value to the concatenated feature tensor to generate an enhanced concatenated feature tensor, and

performing a 2D convolution operation on the enhanced concatenated feature tensor to reduce the number of channels in the enhanced concatenated feature tensor from (N×C) to (N×M), wherein M is less than C; and

encoding the reduced concatenated feature data into a bitstream.

2 . The method of claim 1 , wherein the number of channels, C, is equal to 256 channels and the scale corresponds to a ¼ scale.

3 . The method of claim 2 , wherein M is equal to 32.

4 . The method of claim 1 , wherein encoding the reduced concatenated data into a bitstream includes quantizing and entropy encoding the enhanced concatenated feature tensor including (N×M) channels.

5 . A device comprising one or more processors configured to:

for each of N pictures included in video data, generate feature data including a number of channels, C, corresponding to a scale, such that the generated feature data for each of the N pictures includes a feature tensor including a channel dimension, a height dimension, and a width dimension, (C×H×W);

concatenate the generated feature data about the channel dimension, such that the concatenated feature data includes a concatenated feature tensor including a channel dimension, wherein the number of channels in the concatenated feature tensor is given by N multiplied by the number of channels corresponding to the scale (N×C);

reduce the number of channels in the concatenated feature data to generate reduced concatenated feature data, wherein reducing the number of channels in the concatenated feature tensor includes:

performing a sequence of res2d operations on the concatenated feature tensor to generate a refinement value, wherein each of the res2d operations is a residual block of a compressive autoencoder,

adding the refinement value to the concatenated feature tensor to generate an enhanced concatenated feature tensor, and

performing a 2D convolution operation on the enhanced concatenated feature tensor to reduce the number of channels in the enhanced concatenated feature tensor from (N×C) to (N×M), wherein M is less than C; and

encode the reduced concatenated feature data into a bitstream.

6 . The device of claim 5 , wherein the number of channels, C, is equal to 256 channels and the scale corresponds to a ¼ scale.

7 . The device of claim 6 , wherein M is equal to 32.

8 . The device of claim 5 , wherein to encode the reduced concatenated data into a bitstream includes to quantize and entropy encode the enhanced concatenated feature tensor including (N×M) channels.