A Hybrid Model for Video Compression Based on the Fusion of Feature Compression Framework and Multi-object Tracking Network

This paper proposes a video compression algorithm which bases on the combination of feature compression framework and multi-object tracking network. Here, the multi-object tracking network uses joint detection and embedding (JDE) network. The feature compression framework is consisted of extraction network, encoder/decoder and reconstruction network. More specifically, features of the last layer which is before FPN (Feature Pyramid Network) in JDE are first condensed by extraction network to obtain a low-dimensional compact representation. The compact representation allows for losing feature details, whilst capturing several key important cues within the feature at an extremely low bit-rate cost. Then, the condensed features are encoded and decoded by DCT (Discrete Cosine Transform) method. After that, the features are restored by reconstruction network. Therefore, our method combines deep neural network and DCT, so as to constitute a new hybrid compression architecture to ensure robust and efficient compression. The proposed method is evaluated on object tracking task and the experimental results prove that this proposed framework significantly outperforms both image anchor and feature anchor. In particular, it is reported that 46.70% BD-rate saving can be achieved against the image anchor and 92.17% BD-rate saving can be achieved against the feature anchor for the object tracking task.