Real-VAS: A Real-World Video Amodal Segmentation Dataset

Real-VAS is a zero-shot evaluation dataset for Video Amodal Segmentation, built from composited real-world video. Every amodal ground-truth mask is produced programmatically from an unoccluded source clip, so hidden object boundaries are exact — no human guesswork, no multi-camera 3D rigs, no synthetic appearance gap.

• Project page: https://finlay-hudson.github.io/real_vas
• Code: https://github.com/finlay-hudson/Real-VAS
• Paper: https://arxiv.org/abs/2611.00430

Dataset summary

Clips	300 occlusion + 100 containment (400 total)
Occluded frames	10,640 (72.5% prevalence)
Container types	On-Surface · Articulated · Mobile

Subsets

• occlusion — two-source composites of a moving occluder over a moving occludee. Amodal masks come directly from the unoccluded source clip.
• containment — object entering a snug-fit container. Amodal masks are propagated through the container's tracked rigid/affine transform (validated to mean IoU 97.5%).

Clip layout

occlusion/<clip_name>/
  frames/           # per-frame RGB (JPG)
  modal_masks/      # visible-only object mask (PNG, binary)
  amodal_masks/     # full-object mask incl. hidden regions (PNG, binary)
  annos.json        # per-frame occlusion level + amount
  form.json         # clip-level metadata
containment/<clip_name>/
  (same structure)

annos.json schema:

{"occlusion": [
  {"level": "NO_OCCLUSION" | "SLIGHT_OCCLUSION" | "SEVERE_OCCLUSION" | "NONE",
   "amount": 0.0 - 1.0}
]}

Loading

Annotations only (via datasets)

from datasets import load_dataset
ds = load_dataset("fhudson96/Real-VAS", name="occlusion")  # or "containment"
print(ds["test"][0])   # one row per clip, column "occlusion" = list of per-frame annos

Full clips (frames + masks + annos)

from huggingface_hub import snapshot_download
from pathlib import Path
import json

local_dir = snapshot_download(
    repo_id="fhudson96/Real-VAS",
    repo_type="dataset",
    allow_patterns="occlusion/**",   # drop for everything; swap for "containment/**"
)

clip = Path(local_dir) / "occlusion" / "<clip_name>"
frames  = sorted((clip / "frames").glob("*.png"))
amodal  = sorted((clip / "amodal_masks").glob("*.png"))
modal   = sorted((clip / "modal_masks").glob("*.png"))
annos   = json.loads((clip / "annos.json").read_text())

Source data

Occluder clips for the occlusion subset are drawn from publicly available YouTube videos (source video IDs are visible in clip names) plus in-house capture. All containment clips are in-house captures. Pipeline details — pose estimation, SAM2 masking, CoTracker3 motion filtering, Depth Anything V2 plausibility checks, Generative Omnimatte compositing — are in the paper.

License

Released under CC-BY-NC-4.0. Academic/research use is free with attribution. Commercial use requires written permission from the authors.

Citation

@inproceedings{hudson2026realvas,
  title     = {Real-VAS: A Real-World Video Amodal Segmentation Dataset},
  author    = {Hudson, Finlay G. C. and Smith, William A. P.},
  booktitle = {Under review},
  year      = {2026}
}