Neural Video Compression with Domain Transfer

DCVC-DT is a new neural video compression framework for better performance on unseen video content.

It adapts latent features online during inference and adjusts rate-distortion weighting at the frame level.

Compared with the DCVC-DC baseline, it cut bitrate by up to 6.21% in experiments.

The method targets the common drop in compression quality when test data differs from training data.