DA-SSD: Domain Adaptation for 3D Single Stage Object Detector

Object detection on point cloud is an important task for autonomous driving technology. Long-distance detection is a major problem. Recent researches have demonstrated that good feature representation is the key to 3D object detection, especially for point-based methods. However, due to the physical characteristics of Lidar. Point clouds are densely distributed at short distance and sparsely distributed at long distance, which increases the difficulty for the points-representation learning. In this paper, a simple and effective single-stage detector, named Domain Adaptation for 3D Single Stage Object Detector (DA-SSD), is proposed with a range domain adaptor. Through the range domain adaptor, the knowledge learned on short-distance objects can be transferred to long-distance objects. The problem of uneven point cloud distribution can be alleviated by the proposed module. Extensive experiments show the effectiveness of the proposed DA-SSD.