Multi-frame Context Generation for Neural Video Compression
Qinhao Huang, Zhang Yun, Junle Liu · 2025
Exploring temporal redundancy among frames is the most important task of video compression. However, exist neural video compression methods either only leverage information from a single frame or fuse multiple frames without filtering out uncorrelated information, resulting in suboptimal compression performance. Therefore, we introduce a Multiframe Context Generation (MFCG) module which fuses multiple referenced frames step-by-step by the proposed Conditional Gated Recurrent Unit (CGRU). With the help of intermediate feature of encoder/decoder, MFCG effectively extracts correlated information from referenced frames. Experiments show that the proposed MFCG based Deep Contextual Video Compression (DCVC) achieves an average of $16.6 \%$ bit rate saving comparing with the baseline DCVC, which demonstrates the effectiveness of our proposed MFCG.