- Published on
Perception-Aligned Object Removal Metrics RC-S and RC-T With a Real-World Video Benchmark
- Authors
- Name
- DeepHorus
Xiaomi’s MiLM Plus research team has unveiled PROVE, a new evaluation framework for object removal in videos, set to appear at ACM MM 2026. The system combines two perception-aligned metrics—RC-S for spatial coherence and RC-T for temporal consistency—with PROVE-Bench, a real-world video benchmark.
Object removal models have advanced rapidly, but existing metrics often misjudge their output. Because erasure is an ill-posed, one-to-many task, no single ground truth exists, making PSNR, SSIM, LPIPS, ReMOVE, and CFD unreliable. PROVE addresses this by comparing edited regions locally in deep feature space, using sliding-window Maximum Mean Discrepancy over DINOv2 features, and requires no reference video.
The metrics are implemented in an Apache 2.0 PyTorch repository, needing Python 3.10+ and a single GPU. RC-S runs at 134.6 ms/frame on an RTX 4090, making nightly CI gates practical. Masks are mandatory, and the tool is designed for model bake-offs, inference tuning, training-data filtering, and reward-model shaping.
PROVE targets industries like smartphone gallery apps, short-video editing, e-commerce catalog cleanup, film VFX, and privacy redaction. The authors documented three failure modes of existing metrics: full-reference bias rewards copy-paste, no-reference metrics fail to detect blur, and temporal metrics miss local flicker.
For example, on ROSE-Bench, progressively blurring the masked region did not degrade ReMOVE or CFD, and CFD flagged a correctly restored bicycle seat as a hallucination. Temporal Consistency and Temporal Flickering barely reacted to injected corruptions on DAVIS because they use full-frame features. These results highlight the need for local, perception-aware scoring.
PROVE is not a product feature but an evaluation harness, and it does not handle real-time on-device scoring or large shadows extending beyond the crop. Still, with its local, reference-free design, PROVE offers a practical standard for comparing object-removal models. Acceptance at ACM MM 2026 signals growing recognition of the need for perception-aligned evaluation.
