SPAR-7M: 7 Million Synthetic 3D Scene Annotations

SPAR-7M is a dataset of 7 million synthetic annotations for 3D scenes, created by jasonzhango and hosted on Hugging Face. The dataset was last updated on September 28, 2025. It is derived from third-party datasets including ScanNet and ScanNet++, which must be obtained separately under their respective non-commercial research terms.

Use Cases

• Train 3D object detection models based on synthetic scene annotations.
• Develop semantic segmentation algorithms for point clouds based on labeled 3D scenes.
• Benchmark scene understanding models using a large-scale synthetic annotation corpus.
• Conduct research on domain adaptation between synthetic and real-world 3D data.

Strengths

• Contains 7 million annotations, indicating a large-scale resource.
• Provides synthetic annotations derived from established datasets like ScanNet and ScanNet++.
• Last updated on September 28, 2025, suggesting recent maintenance.

Limitations

• The core image, depth, mesh, and panorama data are not included and must be sourced separately from third parties.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the annotation data itself is unknown, which may limit suitability assessment.

Provenance

Source

Derived from ScanNet and ScanNet++ datasets.

Collection Method

Synthetic annotations created by the author.

Freshness

Last updated 2025-09-28 04:41:22