Skip to content

Diffusion-Vas

Segment hidden objects and fill occluded video regions

Visit official site

What Diffusion-Vas is

The model handles video amodal segmentation and content completion for objects that are partly hidden.

What you can do with it

  • Separate occluded pedestrians or vehicles in surveillance video
  • Repair occluded scene parts in film post-production
  • Read occluded objects in driving scenes

Key features

  • Use conditional generation with context pseudo-depth maps
  • Run a 3D UNet backbone in both stages
  • Fill occluded regions with RGB content

Pros & cons

✓ What we like

  • Handles heavy occlusion and deforming objects
  • Zero-shot transfer from synthetic training to real scenes

! What to watch out for

  • Video data with the target object must be prepared
  • Output masks are compared with real masks for evaluation

FAQ

Who proposed it?

Carnegie Mellon University is named in the material.

What gain is reported?

The material reports up to 13% on occluded amodal segmentation.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review