DeepVCP: An End-to-End Deep Neural Network for Point Cloud Registration using Pytorch

This is a Pytorch implementation of the DeepVCP (Virtual Corresponding Points) network for point cloud registration, based on the original paper by Baidu Research https://songshiyu01.github.io/pdf/DeepVCP_W.Lu_S.Song_ICCV2019.pdf.

The model aligns a source point cloud to a target point cloud by learning the rigid transform (translation + rotation matrix).

Architecture

The architecture consists of

1. Deep Feature Extraction Layer
2. Point Weighting
3. Deep Feature Embedding
4. CPG (Corresponding Point Generation) layer