FVC: A New Framework towards Deep Video Compression in Feature Space

Key Points

motion estimation, motion compression, motion compensation

deformable compensation module

nonlocal attention mechanism

hand-designed module: block-based motion estimation, Discrete Cosine Transform(DCT)

optical flow estimation

complex non-rigid motion pattern

motion information: offset maps for convolution kernels in deformable convolution

auto-encoder style network

end-to-end optimized video compression framework

Proposed Method

Hyperprior entropy model: accurate bitrate estimation

通过CNN估计高斯分布的参数

Quantization:

量化操作不可微

这两个操作和DVC中的操作是一致的

相关论文

Deformable convolutional networks

DVC: An end-to-end deep video compression framework

attention mechanism