SimUDA: Towards Progressive Simulation Data Augmentation in Unsupervised Domain Adaptation for Roadside 3D Object Detection
Wei Gong, Yafei Wang, Bowen Wang, Atong Luo, Shuai Wang, Zexing Li · Advances in transdisciplinary engineering · 2025
Due to different lidar mounting positions and traffic conditions in the real world, the roadside 3D detection model may suffer a great performance degradation when applied to unknown scenarios. However, it is difficult and labor-intensive to reconstruct a highly feature-rich roadside labeled dataset for a new scenario. To enhance the detection accuracy and the robustness of the source pre-trained model when encountering different scenes, this paper proposes SimUDA, a progressive-based domain adaptation method using simulation data augmentation for roadside 3D object detection. Firstly, an incremental roadside simulation data generating and injecting strategy is developed during the source model pre-training stage, where a customized progressive iterative mechanism for data mixing is applied to ensure stability and accuracy convergence of the training model. Then in the transfer learning stage, to avoid the deterioration of generated pseudo labels because of data distortion in the simulation, we iteratively filter out labels with low confidence based on the geometric consistency sequential frames. Compared to the methods without using any simulation data augmentation strategy, our method shows competitive detection performance in both the popular roadside benchmark and the random intersection data collected in the real world.