Paper Project Page Code

Step 1: 3D Scanning. Follow the to record a 3D scan (.ply). Make sure that the z-axis is up (or choose the correct up-axis below) and make sure that the scan is at metric scale, i.e., the point coordinates should be in meters. The .ply should be less than 100 MB, include color and vertex normals. The scene should be metric scale, i.e., the units are in meters. Our model is trained on indoor scenes (appartments, hotels, offices), this is where it will perform best. The performance will likely be reduced on other scenes, for example outdoor scenes.

Instructions Here we provide instructions for recording 3D scans using the "3D Scanner App" for iPhone or iPad with a Lidar sensor (3D meshes recorded with other devices are also welcome). For privacy reasons, please do not record people, remove any personal objects from the scene, and only record on locations where you have permission to do so.

Step 2: 3D Instance Segmentation. Upload your 3D scene (.ply) below to obtain the semantic instance segmentation.

Step 3: Results. For the processed scans, we provide:
- Visualizations
- Predictions (ScanNet FileFormat).

Selected Examples

Selected Failure Cases

Mask3D is currently limited in its ability to identify object classes that were not encountered during training.
Examples of such classes include futons (Japanese beds).