LiFlow: Flow Matching for 3D LiDAR Scene Completion

TL;DR

LiFlow reframes 3D LiDAR scene completion as Flow Matching (FM) to overcome distribution-mismatch issues inherent to diffusion-based methods. By using Nearest Neighbor Flow Matching (NFM) and Chamfer Distance Matching (CDM), LiFlow learns a time-dependent vector field that maps a noisy initial cloud to a complete scene $G$, without assuming Gaussian noise. The method trains a MinkUNet-based vector field with a combined loss $\mathcal{L} = \lambda_{\text{NFM}}\mathcal{L}_{\text{NFM}} + \lambda_{\text{CDM}}\mathcal{L}_{\text{CDM}}$ and performs efficient inference via Euler steps with classifier-free guidance, achieving state-of-the-art CD and JSD on SemanticKITTI and Apollo. While LiFlow excels in global geometric fidelity and occupancy, it shows a slight trade-off in fine-grained 0.1 m voxel IoU, indicating future work to improve high-resolution occupancy without sacrificing global consistency. Overall, LiFlow offers a practically efficient and distribution-consistent alternative to diffusion methods for real-time autonomous driving perception.

Abstract

In autonomous driving scenarios, the collected LiDAR point clouds can be challenged by occlusion and long-range sparsity, limiting the perception of autonomous driving systems. Scene completion methods can infer the missing parts of incomplete 3D LiDAR scenes. Recent methods adopt local point-level denoising diffusion probabilistic models, which require predicting Gaussian noise, leading to a mismatch between training and inference initial distributions. This paper introduces the first flow matching framework for 3D LiDAR scene completion, improving upon diffusion-based methods by ensuring consistent initial distributions between training and inference. The model employs a nearest neighbor flow matching loss and a Chamfer distance loss to enhance both local structure and global coverage in the alignment of point clouds. LiFlow achieves state-of-the-art performance across multiple metrics. Code: https://github.com/matteandre/LiFlow.

Figures (2)

• Figure 2: Qualitative results on a single LiDAR scan from the SemanticKITTI dataset (sequence $08$). Colors depict point height. $\dagger$: with refinement network nunes2024scaling.
• Figure 3: Qualitative results on a single LiDAR scan from the Apollo dataset (sequence $00$). Colors depict point height. $\dagger$: with refinement network nunes2024scaling.