We developed a point cloud annotation tool that enables quick labeling of point clouds with minimal human supervision. Our framework converges to the user desired annotation in an iterative fashion with the user giving a seed for each of the target classes and subsequently giving correction strokes for the segmentations predicted by the network. We tested this tool for part segmentation of point clouds in the ShapeNet dataset and found that we were able to achieve target annotations in significantly lesser number of the clicks as opposed to a manual method. In addition, we also worked on a created a new annotated dataset of reconstructed objects.
For more details, consider having a look at our final presentation and the final report for this project.