Learning-Deep-Learning

PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space

Feb 2019

tl;dr: Use multistage PointNet to enable hierarchical learning.

Overall impression

The paper is an excellent piece of scientific writing full of insightul and concise statements. The paper extends the work of PointNet from multi-stage/hierarchical viewpoint. PointNet++ applies PointNet recursively on a nested partioning of the input data. It is impressive to see the authors uses multi-scale technique to address the loss of robustness in vanilla PointNet++. That said, the idea of using farthest point sampling does not seem to be a “clean” solution. Besides, the computation cost is quite high compared to the vanilla pointnet. See EdgeConv for more details.

Key ideas

Technical details