GPR-Net:Geometric Dynamic Graph Convolutional Neural Network for Low Overlap Point Cloud Registration
Siwen Li, Shihao Xing, Lin Xian Feng, Ziqiang Li, Fancheng Yang, Bowen Deng · 2023
The current point cloud registration methods cannot effectively address low-overlap scenarios. Hence, we present a simple, flexible, and general framework titled GPR-Net for low-overlap point cloud registration. We use Geometric Dynamic Graph Convolutional Neural Network (GeoDGCNN) block to extract the reliable geometric features. Then the cross-attention based on the message passing formulation encodes the two pairs of point clouds for early information exchange, and thus can predict which points lie in the overlap region. Finally, a Random Sample Consensus(RANSAC) algorithm is used to estimate transformation between the source and the reference point cloud. The experimental results show that the registration recall of our method in the 3DLoMatch datasets reaches 67.9%.