Adaptive Graduated Non-Convexity for Point Cloud Registration

Hangqi Ding, Yue Wu, Maoguo Gong, Qiguang Miao, Hao Li, Mingyang Zhang, Wenping Ma

Keywords: Non-Convexity; Point Cloud; Registration

TL;DR: AGNC

Abstract: Point cloud registration is a critical and challenging task in computer vision. It is difficult to avoid poor local minima since the cost function is significantly non-convex. Correspondences tainted by significant or unknown outliers may cause the probability of finding a close-to-true transformation to drop rapidly, leading to point cloud registration failure. Many registration methods avoid local minima by updating the scale parameter of the cost function using graduated non-convexity (GNC). However, the update is usually performed in a fixed manner, resulting in limited accuracy and robustness of registration, and failure to reliably converge to the global minimum. Therefore, we present a novel method to robust point cloud registration based on Adaptive Graduated Non-Convexity (AGNC). By monitoring the positive definiteness of the Hessian of the cost function, the scale in graduated non-convexity is adaptively reduced without the need for a fixed optimization schedule. In addition, a multi-task knowledge sharing mechanism is used to achieve collaborative optimization of non-convex cost functions at different levels to further improve the success rate of point cloud registration under challenging high outlier conditions. Experimental results on simulated and real point cloud registration datasets show that AGNC far outperforms state-of-the-art methods in terms of robustness and accuracy, and can obtain promising registration results even in the case of extreme 99\% outlier rates. To the best of our knowledge, this is the first study that explores point cloud registration considering adaptive graduated non-convexity.

Primary Area: optimization

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8570