Consistent 3D Mask Labeling Made Simple

The 3DIML approach uses a three-step pipeline to convert erratic 2D instance masks into logical 3D instance segmentations. To ensure consistent object classification in spite of segmentation noise, InstanceMap uses a scalable mask association graph based on extended hLoc to associate masks across pictures in the first stage. After resolving ambiguities with a label NeRF, InstanceLift refines these pseudolabels. A quick post-processing step then effectively combines labels that are in disagreement.