Efficient LiDAR Point Cloud Oversegmentation Network

Le Hui, Linghua Tang, Yuchao Dai, Jin Xie, Jian Yang

Abstract

SuperLiDAR is an end-to-end network designed for efficient LiDAR point cloud oversegmentation. By integrating deep metric learning with a range-image-based grouping algorithm, it transforms large-scale point clouds into geometrically and semantically coherent superpoints. SuperLiDAR achieves state-of-the-art accuracy while operating 100x faster than existing methods, facilitating real-time preprocessing for autonomous driving tasks.

Introduction

Oversegmentation in 3D computer vision involves grouping raw points into larger, meaningful clusters known as superpoints. For LiDAR data, this task is particularly challenging due to the sparse, noisy, and irregular nature of the point clouds. Existing approaches often rely on handcrafted features or multi-stage deep learning pipelines, which are either inaccurate or too computationally expensive for real-time autonomous driving applications. SuperLiDAR addresses these challenges by providing a fast, accurate, and end-to-end trainable framework.

Methodology

The SuperLiDAR architecture is composed of three primary stages: low-level point embedding, LiDAR point grouping, and superpoint refinement.

Overview of the SuperLiDAR architecture, including (a) low-level point embedding, (b) LiDAR point grouping via BFS on range images, and (c) superpoint refinement.
  1. Low-Level Point Embedding: A sparse 3D backbone extracts features that are mapped into a low-dimensional embedding space. A local discriminative loss is employed to ensure that points belonging to the same object part have similar embeddings, while distinct parts are pushed apart.
  2. LiDAR Point Grouping: To form clusters efficiently, the network projects sparse 3D points onto a 2D range image. This representation allows a Breadth-First Search (BFS) algorithm to traverse adjacent points in $O(N)$ time. Points are grouped into a superpoint if they satisfy both spatial proximity in 3D and similarity in the learned embedding space.
  3. Superpoint Refinement: To handle points near boundaries that BFS might miss, a refinement module learns the affinity between unassigned points and their $K$-nearest superpoints, ensuring precise boundary delineation.

Experimental Evaluation

The performance of SuperLiDAR was evaluated on two large-scale outdoor datasets:

  • SemanticKITTI: A dataset with 64-beam LiDAR scans across 22 urban sequences.
  • nuScenes: A dataset featuring 32-beam LiDAR scans, characterized by higher sparsity and complexity.

Results and Discussion

SuperLiDAR demonstrates significant improvements in both speed and accuracy compared to prior state-of-the-art methods.

Performance comparison of different methods on the SemanticKITTI validation set across various metrics: Oracle Overall Accuracy (OOA), Boundary Recall (BR), Boundary Precision (BP), and F1 score.
  • Efficiency: SuperLiDAR processes a single scan in approximately 72ms, representing a 100x speedup over previous methods like SPNet. This efficiency makes it suitable for real-time deployment.
  • Accuracy: On SemanticKITTI, the model achieved an Oracle Overall Accuracy (OOA) of 96.21%, outperforming SPNet by approximately 4%.
  • Downstream Utility: Integrating SuperLiDAR superpoints into a standard semantic segmentation network via Multi-Scale Superpoint Aggregation improved the mean Intersection over Union (mIoU) by 2% on SemanticKITTI and 1% on nuScenes.

The use of range images for grouping is a key factor in the model's speed, while deep metric learning allows it to handle complex geometries more effectively than handcrafted features.

Conclusion

SuperLiDAR provides a highly efficient and accurate solution for LiDAR point cloud oversegmentation. While the current grouping stage utilizes a BFS algorithm, future work could explore more natively "deep" grouping mechanisms. Additionally, applying these superpoints to tasks such as 3D object detection and multi-object tracking could further demonstrate the benefits of this adaptive neighborhood representation.

Summary updated

Abstract

Point cloud oversegmentation is a challenging task since it needs to produce perceptually meaningful partitions (i.e., superpoints) of a point cloud. Most existing oversegmentation methods cannot efficiently generate superpoints from large-scale LiDAR point clouds due to complex and inefficient procedures. In this paper, we propose a simple yet efficient end-to-end LiDAR oversegmentation network, which segments superpoints from the LiDAR point cloud by grouping points based on low-level point embeddings. Specifically, we first learn the similarity of points from the constructed local neighborhoods to obtain low-level point embeddings through the local discriminative loss. Then, to generate homogeneous superpoints from the sparse LiDAR point cloud, we propose a LiDAR point grouping algorithm that simultaneously considers the similarity of point embeddings and the Euclidean distance of points in 3D space. Finally, we design a superpoint refinement module for accurately assigning the hard boundary points to the corresponding superpoints. Extensive results on two large-scale outdoor datasets, SemanticKITTI and nuScenes, show that our method achieves a new state-of-the-art in LiDAR oversegmentation. Notably, the inference time of our method is 100× faster than that of other methods. Furthermore, we apply the learned superpoints to the LiDAR semantic segmentation task and the results show that using superpoints can significantly improve the LiDAR semantic segmentation of the baseline network. Code is available at https://github.com/fpthink/SuperLiDAR.
BibTeX
@inproceedings{hui2023efficient,
  author = {Le Hui and Linghua Tang and Yuchao Dai and Jin Xie and Jian Yang},
  title = {{Efficient LiDAR Point Cloud Oversegmentation Network}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2023},
  doi = {10.1109/iccv51070.2023.01650},
}
Figure