A comparative analysis of point sampling strategies in point-based 3D object detection

Point-based 3D object detection has been receiving increasing attention as it can preserve the geometric information of a point cloud and avoid quantization errors or information loss caused by voxelization or projection. Point sampling plays an important role in point-based 3D detectors yet has not been thoroughly explored. In this paper, we conduct a comparative analysis of three point sampling strategies to gain a deep understanding of the effect that each strategy imposes on the final performance and intermediate stages of the network. We introduce density-aware sampling and semantic-aware sampling strategies and fit them into the backbone of a lightweight and effective baseline model, aiming to reduce the density imbalance of the point cloud and better utilize semantic information. The density-aware strategy effectively balances the density but the inference time is not suitable for real-time application. Semantic-aware sampling biased on foreground points achieves a 0.19% improvement over the baseline. Analysis on statistics and visualization reveals future research direction. We build our models on the MMDetection3D platform and evaluate performance on the KITTI dataset.