Combinational Point Sampling for Fast and Accurate On-Device LiDAR 3D Object Detection

While recent LiDARs are evolving to higher distance range and angular resolution, the increase in point rate poses significant challenges in achieving low-latency 3D object detection. Although several recent studies aimed at point sampling for efficient point cloud sampling, they are limited in their application to 3D object detection on outdoor LiDAR scenes due to the high sampling overhead and downstream task accuracy drop. We present Cirrus, an end-to-end system for fast and accurate on-device 3D object detection with a novel combinational point sampling. To enable low-overhead and accuracy-preserving point sampling, we design lightweight sampling methods that effectively leverage the sampling opportunities in outdoor LiDAR point clouds (i.e., sparse object occupancy and redundant short-distance points) and synergistically combine them. Extensive evaluation over various edge devices, 3D object detection models, and datasets show that Cirrus effectively reduces the number of input points by 88%, achieving a 40% reduction in end-to-end latency and 35% decrease in energy consumption without accuracy drop.