Systems and methods for neural-network based video encoding

1. A method comprising:

accessing voxel geometry information for a first frame of a 3D media content;

training a first neural network based on the voxel geometry information of the first frame, such that the first neural network is configured to receive a coordinate of a voxel and output color attributes information for the voxel, wherein the trained first neural network comprises a plurality of weights for each layer of the first neural network;

converting the first plurality of weights of each respective layer of the first neural network into a respective first intermediary matrix of the first frame;

for each respective first intermediary matrix: decomposing the respective first intermediary matrix to create a respective first decomposition data that compromises a plurality of components that approximate the respective first intermediary matrix when combined; and

generating encoding data for the first frame by storing each respective decomposition data for each respective first intermediary matrix;

training a second neural network based on voxel geometry information for a second frame of the 3D media content, wherein the trained second neural network comprises a second plurality of weights for each layer of the second neural network;

converting the second weights of each respective layer of the second neural network into a second respective intermediary matrix;

for each respective second intermediary matrix:

identifying a respective matching first intermediary matrix;

generating a respective second decomposition data with same decomposing as the respective matching first intermediary matrix;

computing respective differences based on the respective second decomposition data and decomposition data of the respective matching first intermediary matrix; and

generating encoding data of the second frame, at least in part, by storing each of the respective differences computed for each of the respective second intermediary matrix.

2. The method of claim 1 , wherein the first neural network comprises a multi-layer fully connected network.

3. The method of claim 2 , wherein the converting the first plurality of weights of each respective layer of the first neural network into the respective first intermediary matrix comprises:

representing the first plurality of weights of the multi-layer fully connected network as a table, where each column of the table represents the first plurality of weights for each respective layer of the first neural network; and

converting each column of the table into a 2-dimensional matrix comprising more than one row and more than one column.

4. The method of claim 1 , wherein the decomposing the respective first intermediary matrix comprises:

performing Singular Value Decomposition (SVD) decomposition of the respective first intermediary matrix to obtain a predetermined number of SVD components.

5. The method of claim 4 , wherein the SVD decomposition is performed to obtain a particular number of SVD components with largest singular values, wherein the number of SVD components is selected to be proportional to the size of the first plurality of weights for each layer of the first neural network.

6. The method of claim 4 , wherein each respective SVD component is represented by a respective singular value, a respective vector with a size equal to a number of rows of the respective first intermediary matrix, and a second respective vector with a size equal to a number of columns of the respective first intermediary matrix.

7. The method of claim 1 , further comprising:

generating a decoded frame by: recovering the first neural network first plurality of weights based on the respective first decomposition data for each respective first intermediary matrix; and recovering color attributes for the decoded frame by inputting voxel geometry information of the first frame into a recovered first neural network that comprises layers with the recovered first neural network first plurality of weights; and

generating for display the decoded frame.

8. The method of claim 1 , wherein the training the second neural network comprises starting the training with the first plurality of weights of the first neural network.

9. The method of claim 1 , wherein the second frame either preceded or follows the first frame in the 3D media content.

10. A system comprising:

control circuitry configured to:

access voxel geometry information for a first frame of a 3D media content;

train a first neural network based on the voxel geometry information of the first frame, such that the first neural network is configured to receive a coordinate of a voxel and output color attributes information for the voxel, wherein the trained first neural network comprises a first plurality of weights for each layer of the first neural network;

convert the first plurality of weights of each respective layer of the first neural network into a respective first intermediary matrix of the first frame;

for each respective first intermediary matrix:

decompose the respective first intermediary matrix to create a respective first decomposition data that compromises a plurality of components that approximate the respective first intermediary matrix when combined; and

generate encoding data for the first frame by storing in a memory circuitry each respective decomposition data for each respective first intermediary matrix; and

memory circuitry configured to store the encoding data;

train a second neural network based on voxel geometry information for a second frame of the 3D media content, wherein the trained second neural network comprises a second plurality of weights for each layer of the second neural network;

convert the second plurality of weights of each respective layer of the second neural network into a second respective intermediary matrix;

for each respective second intermediary matrix:

identify a respective matching first intermediary matrix;

generate a respective second decomposition data with same decomposing as the respective matching first intermediary matrix of the first frame;

compute respective differences based on the respective second decomposition data and decomposition data of the respective matching first intermediary matrix; and

generate encoding data of the second frame, at least in part, by storing each of the respective differences computed for each of the respective second intermediary matrix.

11. The system of claim 10 , wherein the first neural network comprises a multi-layer fully connected network.

12. The system of claim 11 , wherein the control circuitry is configured to convert the first plurality of weights of each respective layer of the first neural network into the respective first intermediary matrix by:

representing the first plurality of weights of the multi-layer fully connected network as a table, where each column of the table represents the first plurality of weights for each respective layer of the first neural network; and

converting each column of the table into a 2-dimensional matrix comprising more than one row and more than one column.

13. The system of claim 10 , wherein the control circuitry is configured to decompose the respective first intermediary matrix by:

performing Singular Value Decomposition (SVD) decomposition of the respective first intermediary matrix to obtain a predetermined number of SVD components.

14. The system of claim 13 , wherein the control circuitry is configured to perform SVD decomposition to obtain a particular a number of SVD components with largest singular values, wherein the number of SVD components is selected to be proportional to the seize of the first plurality of weights for each layer of the first neural network.

15. The system of claim 13 , wherein each respective SVD component is represented by a respective singular value, a respective first vector with a size equal to number of rows of the respective first intermediary matrix, and a second respective vector with a size equal to number of columns of the respective first intermediary matrix.

16. The system of claim 10 , wherein the control circuitry is further configured to:

generate a decoded frame by:

recovering the first neural network first plurality of weights based on the respective first decomposition data for each respective first intermediary matrix; and

recovering color attributes for the decoded frame by inputting voxel geometry information of the first frame into a recovered first neural network that comprises layers with the recovered first neural network first plurality of weights; and

generate for display the decoded frame.

17. The method of claim 1 wherein the first neural network is configured to receive the coordinate of the voxel a three values defining 3D location of the voxel, and output the color attributes information for the voxel as three RGB values.