Weakly supervised semantic segmentation on 3D point clouds

Semantic segmentation on 3D point cloud is a fundamental task to many applications like robotics, augmented reality and self-driving cars. However, the annotation cost on 3D data is high. Weakly supervised learning can reduce the human effort in labeling the data.
Objective: Training point cloud segmentation network with subcloud-level labels (category label for each input sample).
Methodology: We extract the localization information from a point cloud classification network trained with the weak label with multiple attention modules and generate pseudo point-level labels. Then, we train a segmentation network using the pseudo labels.
We achieved competitive results with some fully supervised methods using only subcloud-level labels.
Published in CVPR2020.

For more information, you may contact our professor Yap Kim Hui.