ScanNet
Indoor RGB-D dataset composed of 1513 3D scans annotated voxel by voxel. It covers 20 object classes for 3D semantic segmentation and is designed for interior scene reconstruction, robotic perception, and multi-view deep learning.
1513 annotated RGB-D scans, 2D/3D formats (images, voxels, meshes, depth maps)
MIT
Description
ScanNet is a computer vision dataset dedicated to 3D interior modeling. It contains over 1,500 scenes scanned using RGB-D sensors, each annotated with 3D semantic labels in the form of voxels.
What is this dataset for?
- 3D semantic segmentation
- Reconstruction of interior scenes
- Detecting and tracking objects in closed environments
- Robotic perception and indoor navigation
Can it be enriched or improved?
Yes, the dataset can be combined with other sensors, such as LiDAR or thermal cameras. It is also possible to annotate finer subclasses or to convert it to point cloud format if necessary.
🔎 In summary
🧠 Recommended for
- 3D perception researchers
- Robotic vision
- Reconstructing scenes
🔧 Compatible tools
- PyTorch3D
- Open3D
- MeshLab
- Minkowski Engine
- Kaolin
💡 Tip
Use depth files to create point clouds with RGB alignment for training PointNet or KPConv models.
Frequently Asked Questions
Does ScanNet only provide 2D images?
No, it includes RGB images, depth maps, meshes, and annotated voxelized 3D volumes.
What is the format of annotations for 3D segmentation?
Annotations are made at the voxel level, with 20 classes of semantic objects in the interior scenes.
Can ScanNet be used for robotics projects?
Yes, it is widely used for autonomous navigation, obstacle detection, and the modeling of closed environments.
