Multi-Frame Neural Scene Flow: Learning Bounds and Algorithms

Dongrui Liu, Daqi Liu, Xueqian Li, Sihao Lin, Hongwei Xie, Bing Wang, Xiaojun Chang, Lei Chu

Keywords: Multi-Frame Neural Scene Flow, Spatial and Temporal Feature, Generalization Bound, Large-Scale Point Clouds.

TL;DR: We theotretically analyze the generalization capabilities of NSFP and propose a multi-frame scene flow scheme with bounded generalization error.

Abstract:

Although Neural Scene Flow Prior (NSFP) and its variants have shown remarkable performance in large out-of-distribution autonomous driving, the underlying explanation for their generalization capabilities remains unclear. To this end, we analyze the generalization capabilities of NSFP via uniform stability and find that it exhibits a generalization bound, which is inversely proportional to the number of point clouds. These findings provide solid theoretical evidence to explain the effectiveness of NSFP in large-scale point cloud scene flow estimation tasks for the first time. To enhance practical scene understanding, we extend NSFP and propose a multi-frame neural scene flow (MNSF) scheme, which extracts temporal information across multiple frames. In this way, MNSF has better temporal consistency than NSFP.
Moreover, we theoretically analyze its generalization abilities and demonstrate that it achieves a tight generalization bound with a convergence rate similar to NSFP. Extensive experimental results on large-scale autonomous driving Waymo Open and Argoverse datasets demonstrate that MNSF achieves state-of-the-art performance. The code is attached to the submission.

Primary Area: interpretability and explainable AI

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: 2483