Physics-Aware Diffusion for LiDAR Point Cloud Densification

arXiv cs.CV / 3/31/2026

💬 OpinionSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper addresses LiDAR point cloud densification challenges caused by distance-dependent sparsity in perception tasks, where existing diffusion approaches can produce unrealistic “ghost points” and incur high latency.
It proposes “Scanline-Consistent Range-Aware Diffusion,” reframing densification as probabilistic refinement of a coarse prior rather than free-form generation.
Using Partial Diffusion (SDEdit), the method targets fast inference, achieving high-fidelity densification in about 156ms.
The approach introduces Ray-Consistency loss and Negative Ray Augmentation to enforce sensor/physics constraints and suppress physical hallucinations.
Experiments on KITTI-360 and nuScenes show state-of-the-art performance, improving off-the-shelf 3D detectors without requiring detector retraining, with code planned for release.

Abstract

LiDAR perception is severely limited by the distance-dependent sparsity of distant objects. While diffusion models can recover dense geometry, they suffer from prohibitive latency and physical hallucinations manifesting as ghost points. We propose Scanline-Consistent Range-Aware Diffusion, a framework that treats densification as probabilistic refinement rather than generation. By leveraging Partial Diffusion (SDEdit) on a coarse prior, we achieve high-fidelity results in just 156ms. Our novel Ray-Consistency loss and Negative Ray Augmentation enforce sensor physics to suppress artifacts. Our method achieves state-of-the-art results on KITTI-360 and nuScenes, directly boosting off-the-shelf 3D detectors without retraining. Code will be made available.