CALM-Net: Curvature-Aware LiDAR Point Cloud-based Multi-Branch Neural Network for Vehicle Re-Identification

TL;DR

CALM-Net addresses vehicle Re-Identification from LiDAR point clouds by integrating edge-focused local topology, global point-context via attention, and a learnable curvature embedding to capture local surface variation. The method uses a three-branch fusion with a hybrid training/inference sampling strategy, achieving state-of-the-art mean Re-ID accuracy on nuScenes while maintaining real-time inference. Quantitative results show a ~1.97 percentage-point improvement over strong baselines, with best performance on rigid object classes and some gaps for deformable categories. The curvature-aware representation demonstrates robustness to viewpoint and sparsity, highlighting the practical impact for LiDAR-based autonomous systems.

Abstract

This paper presents CALM-Net, a curvature-aware LiDAR point cloud-based multi-branch neural network for vehicle re-identification. The proposed model addresses the challenge of learning discriminative and complementary features from three-dimensional point clouds to distinguish between vehicles. CALM-Net employs a multi-branch architecture that integrates edge convolution, point attention, and a curvature embedding that characterizes local surface variation in point clouds. By combining these mechanisms, the model learns richer geometric and contextual features that are well suited for the re-identification task. Experimental evaluation on the large-scale nuScenes dataset demonstrates that CALM-Net achieves a mean re-identification accuracy improvement of approximately 1.97\% points compared with the strongest baseline in our study. The results confirms the effectiveness of incorporating curvature information into deep learning architectures and highlight the benefit of multi-branch feature learning for LiDAR point cloud-based vehicle re-identification.

Figures (6)

• Figure -1: The structure of the proposed CALM-Net network. The input LiDAR point cloud is processed through edge convolution and point attention modules, while curvature embedding is encoded from local point eigenvalues. The output features are aggregated into a unified embedding for vehicle Re-ID.
• Figure 0: Visual explanation of curvature information. The relationship between the 3D eigenvalues of a point's local covariance matrix reveals the curvature information of the point surface. The red, blue, and green vectors represent eigenvectors, with their lengths proportional to the corresponding eigenvalues.
• Figure 1: Inaccurate curvature interpretation due to LiDAR sparsity. Due to the fixed angular resolution of LiDAR, certain regions of a 3D object may be misinterpreted with incorrect curvature information.
• Figure 2: Comparison of random and FPS-based subsampling. Random subsampling introduces irregular densities within local patches, whereas FPS produces more uniform coverage across the object surface.
• Figure 3: Sample point clouds of nuScenes dataset object classes. In this dataset, the LiDAR point cloud data has been re-annotated from original nuScenes dataset using the BEVFusion C+L algorithm lidar-reid. The figure displays camera and LiDAR measurements for various object classes, with the point clouds representing the stacked form of multiple time frames.