Semantic segmentation of LIDAR point clouds from the KITTI-360 dataset using a modified PointNet2. This is a Python and PyTorch based implementation in a Jupyter Notebook. The following animation shows the model predictions by color.
This project proposes and implements a new variant of PointNet2 architecture for semantic segmentation of point clouds. The main difference between this implementation and the original one is the methodology for choosing and processing local features. Here I propose an alternative to the farthest point sampling (FPS) algorithm that improves accuracy by roughly 3 percent.
Coming Soon!