LFPS: Learned Farthest Point Sampling

Jonathan Heins, Pascal Kerschke

Keywords: Point Clouds, Farthest Point Sampling, Learned Sampling, Loss Function

TL;DR: We introduce Learned Farthest Point Sampling (LFPS), an innovative approach that combines the advantages of both algorithmic and learned sampling methods.

Abstract:

The processing of point clouds with deep neural networks is relevant for many applications, including remote sensing and autonomous driving with LiDAR sensors. To ensure the computational feasibility of point cloud processing, it is crucial to reduce the cloud's resolution, i.e., its number of points. This downsampling of point clouds requires a deep learning model to abstract information, enabling it to process points within a more holistic context. A traditional technique for reducing the resolution of a point cloud is Farthest Point Sampling (FPS). It achieves a uniform point distribution but does not adapt to the network's learning process. In contrast, learned sampling methods are adaptive to the network but cannot be seamlessly incorporated into diverse network architectures and do not guarantee uniformity. Thus, they can miss informative regions of the point cloud, reducing their effectiveness for large-scale point cloud applications.

To address these limitations and bridge the gap between algorithmic and learned sampling methods, we introduce Learned Farthest Point Sampling (LFPS), an innovative approach that combines the advantages of both algorithmic and learned techniques. Our method relies on a novel loss function designed to enforce a uniform point distribution. We show by theoretical proof that its minima guarantee a uniformity comparable to FPS. Furthermore, we extend the loss function to include information about key points, enabling the network to adaptively influence point selection while preserving uniform distribution in relevant as well as less relevant regions. In experimental studies, we evaluate the performance of LFPS both independently and within existing network architectures. Our results (a) show that LFPS serves as a plug-in alternative for algorithmic sampling methods, particularly as a faster alternative to FPS for large-scale point clouds, and (b) confirm the enhanced performance of LFPS across various tasks, emphasizing its versatility and effectiveness.

Primary Area: learning on graphs and other geometries & topologies

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: 11769