Internet Video Reconstructions for 3D Scene Understanding

SceneVerse++ contains 1019 3D scenes automatically reconstructed from unlabeled internet videos. It was created by bigai to address the scarcity of annotated 3D scene data. The dataset was last updated in April 2026.

Use Cases

• Train instance segmentation models on the automatically generated per-scene segmentation masks.
• Develop 3D scene understanding algorithms using the reconstructed scene geometry and structure.
• Benchmark 3D reconstruction techniques against the 1019 scenes derived from internet videos.
• Pretrain models for downstream tasks like object detection and scene parsing using the unlabeled video data pipeline.

Strengths

• 1019 reconstructed 3D scenes provide a substantial collection for training.
• Data was automatically generated from abundant internet videos, enabling scalable creation.

Limitations

• The dataset size of 1019 scenes may be limited compared to large-scale 2D image datasets.
• Automated generation from internet videos may introduce label noise or reconstruction artifacts.
• Specific details on scene diversity, geographic origin, or video source quality are not provided.

Provenance

Source

bigai on Hugging Face.

Collection Method

Automatically reconstructed from unlabeled internet videos using a designed data engine.

Freshness

Last updated in April 2026.

Geography

Region tag suggests 'us', but specific coverage is unknown.