Leveraging Vehicle-Side Detectors With Graph Priors for Roadside Domain Adaptive 3-D Object Detection

The deployment of LiDAR-based 3D roadside measurements and detections in new scenarios involves a labor-intensive process of data collection and annotation. With the rapid development of the high-quality labeled vehicle-side autonomous driving dataset in last few years, it has potential to directly transfer the knowledge and enhance the robustness of roadside LiDAR-based 3D perception model. However, since the domain gaps between different perspectives, it is still suffering a performance degradation when directly applying the training model to unknown roadside scenarios. To this end, we propose a graph-aware self-supervised domain adaptation (GA-RP3D) method to enhance the roadside 3D perception using vehicle-side 3D object detection models. Firstly, we develop a graph-aware knowledge distillation framework to bridge the domain gap, where the learnable regions of interest (ROIs) with spatiotemporal motion states are used to construct global graph nodes and edges. In the adaptive fine-tuning stage, a decoupled graph consistency alignment loss function with a robust kernel is designed to enhance the quality of graph consistency representation and improve training stability. On popular 3D roadside benchmarks and self-collected roadside data, the proposed method demonstrates significant competitiveness compared with the state-of-the-art in the evaluated settings.