An Investigation of Beam Density on LiDAR Object Detection Performance

TL;DR

This paper addresses beam-density–induced domain gaps in LiDAR-based 3D object detection for autonomous driving. It combines a domain-shift taxonomy with a comprehensive detector-architecture evaluation, followed by density-resampling experiments to isolate beam-density effects and a training-vs-inference domain-gap analysis. A key finding is that a two-stage detector combining voxel- and point-based representations, such as PV-RCNN++, offers the strongest cross-domain robustness, while detectors trained on dense data generalize better for detection tasks; inference shows robustness to moderate beam-density changes, with performance improving mainly due to easier detection rather than better generalization. The work advocates a holistic domain adaptation approach that starts from a robust detector to minimize the initial domain gap, then focuses adaptation efforts on more complex domain shifts, informing practical deployment with real-world beam-density variability.

Abstract

Accurate 3D object detection is a critical component of autonomous driving, enabling vehicles to perceive their surroundings with precision and make informed decisions. LiDAR sensors, widely used for their ability to provide detailed 3D measurements, are key to achieving this capability. However, variations between training and inference data can cause significant performance drops when object detection models are employed in different sensor settings. One critical factor is beam density, as inference on sparse, cost-effective LiDAR sensors is often preferred in real-world applications. Despite previous work addressing the beam-density-induced domain gap, substantial knowledge gaps remain, particularly concerning dense 128-beam sensors in cross-domain scenarios. To gain better understanding of the impact of beam density on domain gaps, we conduct a comprehensive investigation that includes an evaluation of different object detection architectures. Our architecture evaluation reveals that combining voxel- and point-based approaches yields superior cross-domain performance by leveraging the strengths of both representations. Building on these findings, we analyze beam-density-induced domain gaps and argue that these domain gaps must be evaluated in conjunction with other domain shifts. Contrary to conventional beliefs, our experiments reveal that detectors benefit from training on denser data and exhibit robustness to beam density variations during inference.

Figures (8)

• Figure 1: (a) Low-density and (b) high-density scan of vehicles at a similar distance. (c) Overall and beam-density-induced domain gap (in % for IOU=0.4) measured by different methods. The Cross-Domain and Density-Resampling methods fail to assess either the beam-density-induced or overall domain gap, while the Training and Inference Domain Gaps provide a complete picture.
• Figure 2: Comparison of the (a) Truck, (b) Rooftop and (c) Zenseact datasets. Differences in LiDAR beam density are clearly visible. Ground-truth objects are marked by red bounding boxes.
• Figure 3: Average length of vehicles for different countries in the ZOD.
• Figure 4: Comparison of average object sizes for the classes Car, Pedestrian and Cyclist for ZOD (green), Truck dataset (blue) the Rooftop dataset (orange). Object sizes of the Rooftop dataset are significantly larger on average.
• Figure 5: Recording area statistics.