Xiaomi’s MiLM Plus Releases PROVE: Perception-Aligned Object Removal Metrics RC-S and RC-T With a Real-World Video Benchmark
Object removal models are outpacing the metrics that measure them. Xiaomi's new RC-S and RC-T benchmarks fix the ranking problem.
Why it matters
Xiaomi released perception-aligned evaluation metrics (RC-S, RC-T) and a video benchmark for object removal models, addressing a fundamental gap where existing metrics (PSNR, SSIM, LPIPS) frequently misevaluate state-of-the-art diffusion erasers. This matters to practitioners building vision models and to frontier watchers tracking capability measurement improvements.
The key facts
11 to knowExisting metrics (PSNR, SSIM, LPIPS, ReMOVE, CFD) frequently rank object removal outputs incorrectly
New metrics: RC-S and RC-T designed for perception alignment
Real-world video benchmark released by Xiaomi MiLM Plus
Problem: object removal is ill-posed (one-to-many task), no single ground truth
Modern diffusion erasers now reconstruct shadows, reflections, occluded structure convincingly
Xiaomi MiLM Plus released PROVE framework with RC-S and RC-T metrics
Existing metrics (PSNR, SSIM, LPIPS, ReMOVE, CFD) rank object-removal outputs incorrectly
Diffusion-based erasers now reconstruct shadows, reflections, and occluded structure convincingly
Object removal is an ill-posed one-to-many task with no single ground truth
Real-world video benchmark included in release
Perception-aligned metrics designed to match human judgment on removal quality
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Object removal models have improved faster than the metrics used to judge them. Diffusion erasers now reconstruct shadows, reflections and occluded structure convincingly, yet PSNR, SSIM, LPIPS, ReMOVE and CFD frequently rank their outputs the wrong way. The root cause is structural: erasure is an…