Xiaomi's MiLM Plus Releases PROVE: Perception-Aligned Object Removal Metrics RC-S and RC-T with a Real-World Video Benchmark
Xiaomi's MiLM Plus team has released PROVE, a set of perception-aligned metrics for evaluating object removal in videos, along with a real-world benchmark. The metrics, RC-S and RC-T, use local distribution matching on DINOv2 features and require no ground truth. PROVE outperforms existing metrics in correlation with human rankings.
Object removal models have advanced faster than the metrics used to evaluate them, often misranking outputs. Xiaomi's MiLM Plus team has released PROVE, accepted at ACM MM 2026, which includes two metrics: RC-S for spatial coherence and RC-T for temporal consistency, plus PROVE-Bench, a two-tier video benchmark. Both metrics score the edited region locally using sliding-window Maximum Mean Discrepancy on DINOv2 features and require no reference video. PROVE is available as an Apache 2.0 PyTorch repository. The metrics address issues with existing metrics: PSNR, SSIM, and LPIPS are biased towards copy-paste, ReMOVE and CFD reward blur and misfire on hallucinations, and temporal metrics are insensitive to corruption. RC-S achieves 0.59 average Kendall's tau against human rankings, significantly outperforming ReMOVE (0.26) and CFD (0.16). PROVE-Bench includes PROVE-M with 80 paired videos and PROVE-H with 100 hard videos without ground truth. The code and dataset are publicly available, and RC-S runs at 134.6 ms per frame, 13.7x faster than CFD.
- Abbreviations
- PSNR = Peak Signal-to-Noise Ratio
- SSIM = Structural Similarity Index Measure
- LPIPS = Learned Perceptual Image Patch Similarity
- CFD = CLIP-Feature Difference
- MMD = Maximum Mean Discrepancy
- RBF = Radial Basis Function
- CI = Continuous Integration
- GPU = Graphics Processing Unit
Source: MarkTechPost —
original