INOPNet: Ignoring Nonoverlapping Points Network With Initial Rotation Robustness for Partial-to-Partial Point Cloud Registration

Point cloud registration is the process of accurately aligning point clouds captured from different perspectives using a range of 3-D sensors, such as stereo cameras, time-of-flight cameras, and structured light scanners, and is crucial in robotics, 3-D shape measurement, and related fields. In recent years, the field of point cloud registration leveraging deep learning has expanded significantly. Nonetheless, registering point clouds with only partial overlaps presents ongoing challenges. Moreover, the point cloud is often noisy, further complicating practical applications. To address these issues, we propose the ignoring nonoverlapping points network (INOPNet) with initial rotation robustness for partial-to-partial point cloud registration. Our method utilizes a novel feature extractor, the 3-D points convolution (3DPC) module, to obtain high-dimensional, initially rotation-robust hybrid features. An overlapping points prediction (OPP) module is then employed to eliminate nonoverlapping points interference and generate soft correspondences for the source point cloud, effectively converting the partial-to-partial registration into a complete-to-complete registration. Our novel INOPNet framework, extensively tested on ModelNet40, the Stanford 3-D scanning repository, and real-world datasets, surpasses traditional and advanced deep learning-based methods in registering unseen shapes, with isotropic rotation and translation errors of 0.902 and 0.0119, respectively. It also shows robustness against noise, achieving errors of 1.211 and 0.0151 under Gaussian noise for unseen categories. Furthermore, INOPNet strikes a favorable balance between speed and registration accuracy, underscoring its significant potential for applications in robotic grasping and other real-world instrumentation and measurement scenarios. Our source code is available on Github (<uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/vemacular/INOPNet.git</uri>).