Xiaomi’s MiLM Plus Unveils PROVE Object-Removal Benchmark

Video object-removal systems can plausibly fill scenery after deleting a person or item, but the task permits multiple valid outputs, complicating objective scoring. Conventional full-reference measures such as PSNR and SSIM can reward copying rather than genuine removal, while some no-reference methods favor blur and global temporal scores can miss localized defects. Xiaomi’s MiLM Plus team developed PROVE to align image and video object-removal evaluation more closely with human perception, giving model developers a more defensible benchmark for quality control.

The paper was submitted to arXiv on May 14, 2026, and revised on July 30, after Xiaomi Research said on July 14 that PROVE had been accepted by ACM Multimedia 2026. Its RC-S metric uses sliding-window comparisons of DINOv2 patch features to assess spatial blending, while RC-T tracks distributions in shared restored regions across consecutive frames; neither requires a ground-truth restored video. PROVE-Bench pairs an 80-video motion-augmented set with a 100-video difficult set lacking ground truth. Xiaomi Research has released the PyTorch code under Apache 2.0, enabling integration into automated test pipelines.