Efficient Point Cloud Matching for 3D Geometric Shape Assembly

Nahyuk Lee, Juhong Min, Junha Lee, Seungwook Kim, Kanghee Lee, Jaesik Park, Minsu Cho

Primary Area: representation learning for computer vision, audio, language, and other modalities

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.

Keywords: Geometric shape assembly, High-dimensional feature transform, Correlation aggregation, Proxy Match Transform

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Learning to assemble geometric shapes into a larger target structure is a fundamental task with various high-level visual applications. In this work, we frame this problem as geometric registration with extremely low overlap. Our goal is to establish accurate correspondences on the mating surface of the shape fragments to predict their relative rigid transformations for assembly. To this end, we introduce Proxy Match Transform (PMT), an approximate high-order feature transform layer that enables reliable correspondences between dense point clouds of shape fragments, while incurring low costs in memory and compute. In our experiments, we demonstrate that Proxy Match Transform surpasses existing state-of-the-art baselines on a popular geometric shape assembly dataset, while exhibiting higher efficiency than other high-order feature transform methods.

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: 8840