Receptive-field-conforming convolutional models for video coding

A method of encoding a block, comprising:

• obtaining, using a convolutional neural network (CNN), partition decisions for sub-blocks of size (αS)×(αS) of the block, by steps comprising:
• obtaining first feature maps from feature extraction layers of the CNN;
• inputting the first feature maps to a concatenation layer of the CNN, wherein each first feature map of the first feature maps is of a smallest possible partition size S×S of the block, and wherein the concatenation layer receives no feature map that is not of the smallest possible partition size S×S;
• and
• obtaining the partition decisions from classifiers of the CNN, wherein each classifier of the classifiers infers partition decisions for sub-blocks of size (αS)×(αS) of the block for a respective value of α, wherein α is a power of 2;
• and
• encoding the block in a compressed bitstream according to the partition decisions.

An apparatus for decoding a block, comprising:

• a processor configured to:
• decode, from a compressed bitstream, partition decisions for the block;
• and
• decode the block according to the partition decisions, wherein an encoder encoded the partition decisions in the compressed bitstream using a convolutional neural network (CNN) that inferred the partition decisions, the encoder performing:
• obtaining, using the CNN, the partition decisions for sub-blocks of size (αS)×(αS) of the block, by steps comprising:
• obtaining feature maps from feature extraction layers of the CNN;
• inputting the feature maps to a concatenation layer of the CNN, wherein each feature map of the feature maps is of a smallest possible partition size S×S of the block, and wherein the concatenation layer receives no feature map that is not of the smallest possible partition size S×S;
• and
• obtaining the partition decisions from classifiers of the CNN, wherein each classifier of the classifiers infers partition decisions for sub-blocks of size (αS)×(αS) of the block for a respective value of α, wherein α is a power of 2.