Towards Feature Compression of LiDAR Point Cloud for 3D Object Detection

Engineering notes

Experimental results demonstrate that on the KITTI dataset, compared to the MPEG-standardized G-PCC algorithm, our method achieves BD-rate gains of 68.28%, 49.33%, and 70.87% for vehicle, pedestrian, and cyclist detection, respectively, in the object detection task.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Currently, automatic driving represents the primary application area for LiDAR point clouds, with 3D object detection being a crucial task for achieving autonomous navigation. However, the substantial volume of point cloud data poses significant challenges for transmission, necessitating the development of efficient compression algorithms. Most existing point cloud compression methods are optimized for human perception, prioritizing signal fidelity over the requirements of downstream machine tasks. Given that current LiDAR sensing applications typically convert point clouds into 2D Bird’s Eye View (BEV) representations for real-time, high-precision sensing, this paper proposes a novel LiDAR point cloud compression framework specifically designed for 3D object detection. Our framework focuses on compressing the 2D BEV features and is driven by the object detection task. Experimental results demonstrate that on the KITTI dataset, compared to the MPEG-standardized G-PCC algorithm, our method achieves BD-rate gains of 68.28%, 49.33%, and 70.87% for vehicle, pedestrian, and cyclist detection, respectively, in the object detection task.

Links and sources

• Official / arXiv page

Comments