ExtremeOcc-3D is a benchmark introduced with a 2026 ICML Spotlight paper. It contains 264 single-view indoor scenes for evaluating amodal 3D generation under severe object occlusion. Each sample provides rendered images, masks, text labels, and ground-truth 3D assets.
Use Cases
- Benchmarking amodal 3D reconstruction performance based on the provided occluded scene renders and ground-truth assets
- Training generative models to infer complete 3D shapes from occluded views based on the isolated visible object part renders
- Evaluating text-driven 3D generation systems based on the category text labels paired with visual data
- Developing occlusion-aware computer vision algorithms based on the target-object masks
Strengths
- 264 single-view indoor scenes provide a structured benchmark size
- Each sample includes multiple aligned data types: renders, masks, text labels, and 3D assets
- Benchmark is associated with a peer-reviewed ICML 2026 Spotlight paper
Limitations
- Column-level documentation is absent; field semantics must be inferred after download
- Row count is unknown, which may limit suitability assessment
- Description metadata is limited; actual data quality requires manual inspection after download
Provenance
- Source
- KayZhu
- Freshness
- Last updated 2026-05-21 12:21:19; freshness should be verified