3D Instance Segmentation via Multi-Task Metric Learning
Abstract
We proposed a method for 3D instance segmentation of voxel-based scenes. Our approach is based on metric learning and the first part assigns all voxels belonging to the same object instance feature vectors that are in close vicinity. Conversely, voxels belonging to different object instances are assigned features that are further apart from each other in the feature space. The second part estimates directional information of object centers, which is used to score the segmentation results generated by the first part.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 27, 2019
- Accession Number
- AD1141153
Entities
People
- Bernard Ghanem
- Jean Lahoud
- Marc Pollefeys
- Martin R. Oswald
Organizations
- ETH Zurich
- King Abdullah University of Science and Technology