Xiaomi Open-Sources CVPR 2026 Winning Tech SVOR to Efficiently Remove Passersby from Videos

Xiaomi recently announced the open-source of SVOR, a video object removal technology that can efficiently remove passersby from videos.

Existing video removal methods often face challenges in real-world scenarios: inaccurate mask edges identified by models, residual shadows after object removal, and flickering when processing fast-moving targets frame by frame.

The SVOR framework addresses these issues specifically: it reduces motion jitter via the MUSE windowed joint strategy, improves mask tolerance with DA-Seg denoising-aware segmentation, and handles shadow and reflection residues through two-stage curriculum training. This technology won the CVPR 2026 Physical Perception Video Instance Removal Challenge, and related papers and code have been open-sourced.