Learning from partially labeled 3D point clouds more efficiently

Annotating 3D LiDAR point clouds for semantic segmentation requires extensive manual effort, making semi-supervised learning essential for practical deployment. CoLLiS addresses the confirmation bias problem in existing semi-supervised LiDAR methods by training multiple representations simultaneously as coequal students, each learning from the others while monitoring disagreements to resolve contradictory predictions. Testing on three datasets shows consistent improvements over prior methods, with particularly strong gains when labels are scarce. The approach integrates pseudo-labeling and training in a single step rather than the two-step pipeline used by competing methods.