Deep Parametric Continuous Convolutional Neural Networks

July 2019

tl;dr: Generalization of 2d conv to point cloud. Similar idea to PointNet and PointNet++, and Point CNN.

Overall impression

From Raquel’s group. This paper is similar to pointNet in that it aggregates and propagates information directly on point cloud. However one of the advantage is that the output may be different from the input. So it could be a fancy interpolation method.

Cons: The math is not rigorous at all. Especially for Fig. 2, the dimensions in the architecture is completely wrong and misleading.

The method is used in ContFuse paper to merge features from camera and lidar. ContFuse used MLP to output results directly instead of modeling the weights to sum over features.

Key ideas

Technical details