Surface Representation in LiDAR Scenes

Haoxi Ran, Xiangru Huang, Vitor Campagnolo Guizilini, Yue Wang

Primary Area: applications to robotics, autonomy, planning

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: Surface Representation, LiDAR Segmentation, Representation Learning, Autonomous Driving

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

TL;DR: We propose a novel framework for LiDAR segementation that effectively processes LiDAR point clouds for autonomous driving scenarios, addressing challenges posed by sparse, uneven, and large-scale data.

Abstract: Learning from point clouds entails knowledge of local shape geometry. Recent efforts have succeeded in representing synthetic point clouds as surfels. However, these methods struggle to deal with LiDAR point clouds captured from real scans, which are sparse, uneven, and larger-scale. In this paper, we introduce \textbf{RealSurf}, a general framework that processes point clouds under extreme conditions like autonomous driving scenarios. We identify several key challenges in applying surface representations to real scans and propose solutions to these challenges: Point Sliding Module that jitters point coordinates within the reconstructed surfels for geometric feature computation, and LiDAR-based surfel reconstruction process that enables models to directly construct surfels from LiDAR point clouds by attenuating unevenness. We evaluate our approach on a diverse set of benchmarks, including nuScenes, SemanticKITTI, and Waymo. RealSurf, with a simple PointNet++ backbone, outperforms its counterparts by a significant margin while remaining efficient. By achieving state-of-the-art results on three benchmarks through a fair and unbiased comparison, RealSurf brings renewed attention to the effectiveness of point-based methods in LiDAR segmentation. Code will be publicly available upon publication.

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: 1501