OpenScan is a benchmark for generalized open-vocabulary 3D scene understanding. It extends ScanNet200-style scene understanding with eight object-attribute aspects: material, affordance, property, type, manner, synonyms, requirement, and element. This dataset mirror contains validation annotations and evaluation ground-truth files for a AAAI 2026 paper.
Use Cases
- Benchmarking 3D scene understanding models based on eight object-attribute aspects.
- Evaluating model performance on open-vocabulary tasks using validation annotations.
- Training models for generalized scene understanding based on the ScanNet200-style framework.
Strengths
- Focuses on eight specific object-attribute aspects for detailed scene analysis.
- Provides validation annotations and ground-truth files for a published AAAI 2026 paper.
Limitations
- Description metadata is limited; actual data quality requires manual inspection after download.
- Column-level documentation is absent; field semantics must be inferred after download.
Provenance
- Source
- huggingface
- Freshness
- Last updated 2026-06-04 16:48:07; freshness should be verified.