RID-Net: A Hybrid MLP-Transformer Network for Robust Point Cloud Registration

Abstract

The robustness of correspondence-based point cloud registration relies on transformation invariance and intrinsic distinctiveness of the descriptors computed for registration. However, for challenging scenarios with different objects having similar local geometry and low point cloud overlap, existing descriptors struggle to retain these properties. This inevitably leads to a considerable loss in matching inlier rates and compromised registration. To address the issue, we propose RID-Net that computes robust Rotation-Invariant and Distinctive descriptors for point cloud registration. Our model works on the philosophy of locally accurate feature abstraction for the points while also accounting for the surroundings. Specifically, the locally accurate representation generated by our rotation-invariant MLP-based module is further enhanced through a proposed spherical shell attention mechanism to explicitly focus on the surrounding geometry. The overall network is designed by staking the proposed component, while eventually leveraging coarse-to-fine correspondences and a robust pose estimator to compute the ultimate output. With extensive evaluation on indoor, outdoor and synthetic point cloud data, we establish the efficacy of the proposed RID-Net. In particular, our method improves the inlier ratio by about 4% on 3DMatch/3DLoMatch dataset and reduces the rotation and translation errors by 26.1% and 19.7% on the KITTI dataset.