This repository contains a PyTorch implementation of PointNet for point cloud classification and segmentation.
Overview PointNet is a pioneering deep learning architecture that directly operates on point clouds by learning point-wise features and combining them through max pooling operations. This allows the network to be invariant to input permutation.
Key components implemented:
PointNetEncoder - Encodes point features through linear layers
PointNetModule - Combines global and local point features
PointNetFull - Implements a T-Net to learn input transforms
IoU - Computes per-class and mean IoU for evaluation
KITTI dataset loading and preprocessing
PyTorch
NumPy
OS file operations
title={Pointnet: Deep learning on point sets for 3d classification and segmentation},
author={Qi, Charles R and Su, Hao and Mo, Kaichun and Guibas, Leonidas J},
journal={Proc. Computer Vision and Pattern Recognition (CVPR), IEEE},
year={2017}
}