Unsupervised Domain Adaptation for Monocular 3D Object Detection via Self-training

Monocular 3D object detection (Mono3D) has achieved unprecedented success with the advent of deep learning techniques and emerging large-scale autonomous driving datasets. However, drastic performance degradation remains an unwell-studied challenge for practical cross-domain deployment as lack labels on target domain. In this paper, we first comprehensively investigate significant underlying factor domain gap in Mono3D, where critical observation is a depth-shift issue caused by geometric misalignment domains. Then, propose STMono3D, new self-teaching framework unsupervised adaptation Mono3D. To mitigate depth-shift, introduce geometry-aligned multi-scale training strategy to disentangle camera parameters guarantee geometry consistency Based this, develop teacher-student paradigm generate adaptive pseudo Benefiting from end-to-end that provides richer information labels, quality-aware supervision take instance-level confidences into account improve effectiveness target-domain process. Moreover, positive focusing dynamic threshold are proposed handle tremendous FN FP samples. STMono3D achieves remarkable all evaluated datasets even surpasses fully supervised results KITTI dataset. best our knowledge, study explore effective UDA methods