PCB-RandNet: Rethinking Random Sampling for LIDAR Semantic Segmentation in Autonomous Driving Scene

TL;DR

PCB-RandNet tackles the bias inherent in Random Sampling for outdoor LiDAR semantic segmentation by introducing Polar Cylinder Balanced Random Sampling to evenly represent points across distance ranges. It adds a Sampling Consistency Loss to stabilize learning when using different sampling strategies, optimized with learnable uncertainty weights. Empirical results on SemanticKITTI and SemanticPOSS show consistent improvements in mIoU, demonstrating the practicality of distance-aware sampling for large-scale outdoor scenes. The approach maintains inference efficiency while enhancing learning dynamics, and the code is released for public use.

Abstract

Fast and efficient semantic segmentation of large-scale LiDAR point clouds is a fundamental problem in autonomous driving. To achieve this goal, the existing point-based methods mainly choose to adopt Random Sampling strategy to process large-scale point clouds. However, our quantative and qualitative studies have found that Random Sampling may be less suitable for the autonomous driving scenario, since the LiDAR points follow an uneven or even long-tailed distribution across the space, which prevents the model from capturing sufficient information from points in different distance ranges and reduces the model's learning capability. To alleviate this problem, we propose a new Polar Cylinder Balanced Random Sampling method that enables the downsampled point clouds to maintain a more balanced distribution and improve the segmentation performance under different spatial distributions. In addition, a sampling consistency loss is introduced to further improve the segmentation performance and reduce the model's variance under different sampling methods. Extensive experiments confirm that our approach produces excellent performance on both SemanticKITTI and SemanticPOSS benchmarks, achieving a 2.8% and 4.0% improvement, respectively. The source code is available at https://github.com/huixiancheng/PCB-RandNet.

Figures (4)

• Figure 1: (a): Visualization of one frame of LiDAR scan on the SemanticKITTI dataset. (b):Visualization of the corresponding point cloud scan distance distribution. (c): Visualization of the distance distribution of all training data. (d): Segmentation performance of the baseline method and our method at different distances.
• Figure 2: Overall pipeline of PCB-RandNet. Our network structure remains the same as RandLA-Net, with only changes to the first sampling method. Two different class prediction distributions $P_1$ and $P_2$ can be obtained by forwarding the input point cloud twice by models with different sampling methods but shared parameters. Then, the sampling consistency loss between the two class prediction probabilities is calculated and used to supervise the model. $(N,D)$ represents the number of points and feature dimension respectively. RS: Random Sampling, PCB-RS: Polar Cylinder Balanced Random Sampling, FC: Fully Connected layer, LFA: Local Feature Aggregation, MLP: shared Multi-Layer Perceptron, US: Up-sampling, DP: Dropout. (Please note that as shown by the dashed line, the upper branch and sampling consistency loss is only used during training and discarded during inference, i.e., it is disposable. Therefore, it does not introduce any computational burden during inference.)
• Figure 3: Comparison of qualitative visualization and quantitative statistical analysis between RS and PCB-RS. The upper part shows the visualization of the point cloud, and the lower part shows the distribution of that point cloud at different distance ranges. The subplot on the right shows the distribution of the number of points within the cylindrical block when using the PCB-RS method. Best viewed in color and zoomed in for more detail.
• Figure : Simple example algorithm of RS and PCB-RS