Network and system for neural pose transfer

A network for neural pose transfer, comprising a pose feature extractor for receiving an input pose, and a style transfer decoder for receiving for extract pose feature, wherein:

• the pose feature extractor comprises a first number of convolution layers and a first number of Instance Norm layers;
• the first number of convolution layers and the first number of Instance Norm layers are iteratively and serially connected;
• the style transfer decoder comprises a second number of convolution layers, a third number of SPAdaIn residual blocks, and a tanh layer;
• the second number is one more than the third number;
• the second number of convolution layers and the third number of SPAdaIn residual blocks are iteratively and serially connected, with the convolution layers sandwiching the SPAdaIn residual blocks;
• and the tan h layer is the last layer of the style transfer decoder;
• and
• an input identity mesh is concatenated with the output of the pose feature extractor and then input to the style transfer decoder, and the input identity mesh is meanwhile fed to each SPAdaIn residual block of the style transfer decoder.

A system for neural pose transfer, comprising an input device, a processor for processing input data, and an output device for outputting processed data;

• wherein the processor is configured to build a computing model comprising:
• a pose feature extractor for receiving an input pose, and a style transfer decoder for receiving for extract pose feature, wherein:
• the pose feature extractor comprises a first number of convolution layers and a first number of Instance Norm layers;
• the first number of convolution layers and the first number of Instance Norm layers are iteratively and serially connected;
• the style transfer decoder comprises a second number of convolution layers, a third number of SPAdaIn residual blocks, and a tanh layer;
• the second number is one more than the third number;
• the second number of convolution layers and the third number of SPAdaIn residual blocks are iteratively and serially connected, with the convolution layers sandwiching the SPAdaIn residual blocks;
• and the tanh layer is the last layer of the style transfer decoder;
• and
• an input identity mesh is concatenated with the output of the pose feature extractor and then input to the style transfer decoder, and the input identity mesh is meanwhile fed to each SPAdaIn residual block of the style transfer decoder.