WIN: Variable-View Implicit LIDAR Upsampling Network
Keywords: LiDAR Upsampling, Deep Learning, Autonomous vehicle system
Abstract: LiDAR upsampling aims to increase the resolution of sparse point sets obtained from low-cost sensors, providing better performance for various downstream tasks. Most existing methods transform LiDAR points into range view and design complex neighborhood point interpolation strategies to improve the resolution of point clouds. However, they overlook that the range image representation is insufficient to describe complex local geometric relationships, which limits the geometric accuracy of upsampled points.
To address this issue, we propose WIN, a Variable-View Implicit Network.
First, we decouple the range image into two novel virtual view representations to compensate for the missing geometric information during range view-based interpolation. Secondly, to fuse the interpolation results of different views, we model the fusion process as a probability distribution problem instead of a simple binary classification task. We introduce a contrast selection module, which captures the feature differences between two representations and outputs the view confidence score for each upsampled point. The underlying idea is that the complementarity of the information is proportional to the feature difference between the two views. Motivated by this insight, we design a loss function based on probabilistic modeling to supervise the results of the selection module.
As a result, compared with the current state-of-the-art (SOTA) method ILN, WIN introduces a small number of parameters (+0.4M) but achieves a +4.5\% increase in the MAE metric on the CARLA dataset. Furthermore, our method outperforms all existing methods in a downstream task (Depth Completion). The pre-trained model and code will be released upon acceptance.
Primary Area: applications to 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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
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: 2439
Loading