ErenAta00/VOID-Quadmask-Dataset

VOID-Compatible Quadmask Counterfactual Video Dataset

The first publicly available, pre-built quadmask-annotated counterfactual video dataset for physics-aware video object removal, inspired by and fully compatible with Netflix/VOID (arXiv:2604.02296).

---

What is this?

VOID introduced a powerful framework for removing objects from videos while correcting downstream physical interactions. Their key innovation is the quadmask — a 4-value segmentation mask that tells the model which regions need physical correction after object removal.

While VOID released their model weights and generation code, they did not release pre-built training data. This dataset fills that gap by providing 200 ready-to-use counterfactual scene pairs with full quadmask annotations.

---

Dataset Structure

Each scene contains a counterfactual pair: the original video (with all objects) and the counterfactual video (with the target object removed and physics re-simulated).

scene_XXXX/
├── rgb_full.avi          # V: Original video (all objects present)
├── rgb_removed.avi       # V̂: Counterfactual video (target removed, physics re-run)
├── mask_lossless.avi     # Mq: Quadmask video (FFV1 lossless codec)
├── mask.mp4              # Mq: Quadmask video (MP4, quantized to exact values)
├── metadata.json         # Scene metadata and pipeline configuration
└── quadmask_frames/      # Individual quadmask frames as PNG
    ├── quadmask_0000.png
    ├── quadmask_0001.png
    └── ...

Quadmask Values (VOID-compatible)

Value	Color	Meaning
0	⬛ Black	Object region — pixels belonging to the removed object
63	🔲 Dark gray	Overlap — object pixels that also show physical interaction
127	⬜ Light gray	Affected area — regions where physics changed due to removal
255	🟩 White	Background — unchanged regions

---

Technical Specifications

Property	Value
Number of scenes	200
Frames per scene	216 (9 seconds)
Resolution	672 × 384
FPS	24
Total frames	43,200
Mask codec	FFV1 lossless (AVI) + quantized MP4
Mask cleanliness	100% — all frames contain only [0, 63, 127, 255]

---

Interaction Categories

This dataset includes 12 structured interaction categories, several of which go beyond the random-drop scenarios in Kubric:

Category	Description
Stack Collapse	Block tower collapses when middle block removed
Domino Chain	Chain reaction stops when first domino removed
Ramp Collision	Rolling object removed, target objects stay still
Bowling Pins	Bowling ball removed, pins remain standing
Platform Drop	Platform removed, objects fall
Newton's Cradle	Pendulum ball removed, momentum transfer stops
Balance Board	Weight removed from seesaw, board tips opposite
Cascading Shelves	Bottom shelf removed, upper shelves collapse
Chain Break	Chain link removed, lower segment falls
Support Removal	Arch support removed, beam collapses
Projectile Impact	Projectile removed, wall remains intact
Multi-Object Push	Pusher removed, objects stay in place

---

⚙️ Quadmask Generation Pipeline

The affected-area masks are computed using a 4-method GPU ensemble:

Method	Weight	Role
RAFT Optical Flow	35%	Motion difference between V and V̂
DINOv2 Semantic Diff	35%	Structural/semantic feature differences
SSIM (GPU)	15%	Perceptual quality differences
RGB Pixel Diff	15%	Raw pixel-level differences

• Post-processing includes morphological cleanup, gridification (16px cells), and temporal smoothing (window=3).
• Object mask (Mo) is derived from the V − V̂ difference in the first frame, before physics diverges — the only difference at t=0 is the removed object itself.

Generation Stack

• Simulation: Unity 6 HDRP with deterministic PhysX
• Quadmask Pipeline: Python + PyTorch on Google Colab
• GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition (102GB VRAM)
• Processing speed: 62 fps (v3.2), 200 scenes in 27 minutes
• Success rate: 200/200 (0% failure)

---

Usage

from huggingface_hub import hf_hub_download
import cv2
import json

# Download a scene
scene = "scene_0000"
for f in ["rgb_full.avi", "rgb_removed.avi", "mask_lossless.avi", "metadata.json"]:
    hf_hub_download(
        repo_id="ErenAta00/VOID-Quadmask-Dataset",
        filename=f"{scene}/{f}",
        repo_type="dataset",
        local_dir="./void_data"
    )

# Read quadmask
cap = cv2.VideoCapture(f"./void_data/{scene}/mask_lossless.avi")
ret, frame = cap.read()
mask = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
# mask values: 0 (object), 63 (overlap), 127 (affected), 255 (background)

---

Comparison with VOID Training Data

Feature	VOID (Kubric)	VOID (HUMOTO)	This Dataset
Scenes	~1,900	~4,500	200 (prototype)
Source	Blender/PyBullet	MoCap	Unity 6 HDRP
Interaction types	Random drops	Human-object	12 structured categories
Joint/mechanical	❌	❌	✅ (Newton's cradle, seesaw, chain)
Pre-built data	❌ (code only)	❌ (code only)	✅ (ready to use)
Mask format	MP4 (lossy)	MP4 (lossy)	FFV1 lossless + MP4
Open access	Code only	Code only	Full dataset

---

Citation

If you use this dataset, please cite both this dataset and the original VOID paper:

@misc{ata2026void_quadmask_dataset,
  title={VOID-Compatible Quadmask Counterfactual Video Dataset},
  author={Eren Ata},
  year={2026},
  url={[https://huggingface.co/datasets/ErenAta00/VOID-Quadmask-Dataset](https://huggingface.co/datasets/ErenAta00/VOID-Quadmask-Dataset)}
}

@misc{motamed2026void,
  title={VOID: Video Object and Interaction Deletion},
  author={Saman Motamed and William Harvey and Benjamin Klein and Luc Van Gool and Zhuoning Yuan and Ta-Ying Cheng},
  year={2026},
  eprint={2604.02296},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

License

Apache 2.0

🙏 Acknowledgments

This dataset was inspired by and built to be compatible with VOID by Netflix and INSAIT. We thank the VOID team for making their model and code openly available.