Dataset · manycore-research
3D point cloud test set for spatial LLMs
manycore-research/SpatialLM-Testset
107 noisy 3D point clouds reconstructed from regular RGB video, with room layout annotations.
What's inside
A small evaluation dataset containing 107 preprocessed point clouds and ground truth (GT) spatial layout annotations. Unlike clean scans from RGBD cameras, these 3D scenes are reconstructed from monocular RGB video using the MASt3R-SLAM algorithm, and therefore contain characteristic noise and occlusions.
What you can do with it
- Evaluating the robustness of 3D-VLMs and spatial models to noise when reconstructing scenes from regular video
- Testing room layout recognition algorithms on incomplete point clouds
- Size
- up to 1K rows
- Downloads this month
- 2.6K
- Likes
- 60
Models trained on it
How to load it
from datasets import load_dataset
ds = load_dataset("manycore-research/SpatialLM-Testset")