SOCR: Simultaneous Overlap Prediction and Correspondence Estimation for Point Cloud Registration in Real-Time

Point cloud registration is a crucial task in 3D computer vision. Correspondence-based methods highly rely on the quality of matching points. However, existing methods still suffer from low efficiency, precision, and recall. This letter introduces SOCR, a method that can perform overlap prediction and correspondence estimation simultaneously in real-time. To reject non-matchable points, SOCR focuses on overlap regions and extracts reliable overlap-correspondence. We propose a dual branch transformer to learn feature descriptors that are conditioned on both point clouds. A spatial-aware loss is also introduced to address the problem of remote mismatching and to encourage SOCR to rapidly focus on the target areas. The experiments on indoor and outdoor datasets show that SOCR is superior to other existing methods. Owing to its independence from key points and RANSAC, the proposed method achieves a matching performance of 20 FPS on a single 1080Ti GPU. SOCR can be integrated into Lidar SLAM as a powerful pose estimator for further application.

I. Introduction

With the rapid development of 3D sensors, point cloud has become the primary data format for recording 3D scenes [1]. Point cloud registration is a fundamental task in the areas of 3D reconstruction [2], autonomous driving [3], [4], and augmented reality [5]. It aims to estimate the transformation matrix that aligns two partially overlapping point clouds. In reality, noise and complex scenarios can make the task extremely challenging.