SpatialLM: 12,328 Indoor Scenes and 54,778 Annotated Rooms

SpatialLM provides point clouds and 3D annotations for 12,328 indoor scenes and 54,778 rooms. Created by manycore-research in 2025, this synthetic dataset was developed by professional 3D designers for production-level indoor scene understanding.

Use Cases

• Training 3D perception models using the ground-truth 3D annotations
• Indoor scene segmentation using point cloud data from 54,778 rooms
• Benchmarking spatial reasoning in Large Language Models using the paired scene-text data

Strengths

• Includes 54,778 individual rooms with ground-truth 3D annotations
• Professional-grade synthetic quality designed for real-world production
• Large-scale diversity across 12,328 indoor scenes

Limitations

• Synthetic data origin may not fully capture the noise characteristics of physical LiDAR or RGB-D sensors
• Restricted to non-commercial use under the CC BY-NC 4.0 license

Provenance

Source

manycore-research

Collection Method

synthetic

Geography

United States