TransVFC: A transformable video feature compression framework for machines

Yuxiao Sun, Yao Zhao, Meiqin Liu, Chao Yao, Huihui Bai, Chunyu Lin, Weisi Lin

Abstract: Highlights•This paper presents a Transformable Video Feature Compression (TransVFC) framework for the field of Video Coding for Machines (VCM). The framework offers a “compress-then-transfer” solution for video feature compression in multitask scenarios.•This paper introduces an innovative neural-based video feature codec to squeeze redundancy in the feature domain. It contains a scheme-based inter-prediction module and a perception-guided conditional coding module for temporal and spatial redundancy removal.•This paper proposes a lightweight feature space transform module to efficiently transfer intermediate features to various downstream tasks without the need for retraining or redeploying the upstream feature codec and downstream networks. This greatly enhances the scalability and flexibility of the feature compression framework across multiple downstream tasks.

External IDs:dblp:journals/pr/SunZLYBLL26