SP-MAE: Lidar Beam-Induced Domain Adaptation for 3D Object Detection via Semantic Point Masked Autoencoders

• Published in: 2023 3rd International Conference on Computer, Control and Robotics (ICCCR) pp. 5 - 11
• Main Authors: Liang, Yaolin, Tu, Zhiming, Wu, Tao
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.03.2023
• Subjects: 3D object detection; Adaptation models; Laser radar; Masked autoencoders; Performance evaluation; Point cloud compression; Semantics; Three-dimensional displays; Training; Unsupervised domain adaptation
• DOI: 10.1109/ICCCR56747.2023.10193909

In autonomous driving scenarios, lidar-based object detectors are widely used in 3D object detection. For the sake of safety redundancy, some self-driving vehicles will be equipped with multiple lidars with different beams, but training multiple lidar-based object detectors requires high costs. In our experiments, the performance of a lidar-based detector model trained on a high-beam lidar domain drops dramatically when transferred to a low-beam lidar domain. At present, there is no public 3D object detection dataset across-lidar-beams domains. In this paper, a 3D object detection dataset adapted across-lidar-beams domain is produced by the LGSVL simulator. The reason for the performance degradation of the lidar-based detectors across lidar-beams domains are mainly due to the large gap in the foreground object point count. To solve this problem, this paper proposed a semantic point generation method based on masked autoencoders, which can bridge the data gap between lidar beam-induced domains by generating more foreground semantic points so as to realize the migration of the detector model. In addition to solving the problem of lidar beam-induced domains migration, SP-MAE can also improve the performance of 3D lidar-based detectors. Experiments showed that the PVRCNN detector can improve 5.11 3D AP of pedestrian on the KITTI dataset after using SP-MAE.