HorizonRobotics introduced ARSG-110K, a large-scale scene-level dataset comprising over 110,000 diverse scenes and 3 million annotated images. It provides high-fidelity 3D ground truth, including accurate object-level data, layout, and annotations. The dataset was last updated on April 8, 2026.
Use Cases
- Training compositional 3D scene generation models based on the provided 3D object-level ground truth.
- Evaluating in-place scene completion models based on the dataset's high-fidelity 3D scene data.
- Developing models for scene layout understanding based on the annotated layout information.
- Benchmarking 3D scene understanding algorithms based on the large-scale, diverse scene collection.
Strengths
- Contains over 110,000 diverse scenes, providing substantial scale.
- Includes 3 million annotated images for detailed scene representation.
- Provides high-fidelity 3D ground truth, including object-level data and layout.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count for individual data points is unknown, which may limit suitability assessment.
- Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
- Source
- HorizonRobotics
- Collection Method
- Introduced as part of the paper '3D-Fixer: Coarse-to-Fine In-place Completion…'; method likely involves 3D scene capture and annotation.
- Time Range
- null
- Freshness
- Last updated 2026-04-08 04:22:06; freshness should be verified.
- Geography
- null