MegaUnScene is a benchmark dataset of Internet scenes unseen by existing 3D foundation models. It was created by researchers from Cornell University, New York University, and Harvard University's Kempner Institute. The dataset includes three test splits for two evaluation tasks, including relative pose estimation.
Use Cases
- Benchmarking 3D foundation models based on unseen Internet scenes.
- Evaluating relative pose estimation performance based on the UnScenePairs test split.
- Testing model generalization on novel, out-of-distribution 3D scenes.
Strengths
- Created by a multi-institution team from Cornell University, New York University, and Harvard University.
- Specifically designed for benchmarking 3D foundation models on unseen data.
- Includes multiple test splits for at least two distinct evaluation tasks.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count, file formats, and license are unknown, which may limit suitability assessment.
- Data may reflect source bias inherent to the Internet scenes used for collection.
Provenance
- Source
- Cornell University, New York University, Harvard University (Kempner Institute).
- Collection Method
- Likely collected from Internet scenes.
- Time Range
- null
- Freshness
- Last updated 2026-02-26 20:52:11; freshness should be verified.
- Geography
- null