Reproduce “Prediction and Reference Quality Adaptation for Learned Video Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xihua Sheng, Li Li, Dong Liu, Houqiang Li
Researchers at the University of Science and Technology of China introduce Prediction Quality Adaptation (PQA) and Reference Quality Adaptation (RQA) modules into a learned video codec, alongside a repeat-long training strategy, to enhance compression efficiency and reduce error propagation. This approach achieves a -12.2% BD-rate saving against DCVC-FM on the HEVC Class B dataset (RGB PSNR) and produces visually clearer reconstructions with minimal increases in computational complexity.
Key insights
- Explicitly discerning and adapting to spatial and channel-wise prediction quality differences within temporal contexts can significantly improve the utilization of temporal information and enhance compression.
- Dynamic adaptation of transform networks through spatially variant filters, based on reference frame quality, is an effective mechanism to stabilize reconstruction quality and reduce error propagation in learned codecs.
- Combining a strategy to generate diverse reference qualities with extended cascaded training sequences enables more robust learning for reference quality adaptation, mitigating the accumulation of reconstruction errors over time without excessive resource demands.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.