| Summary: | <b>Unsupervised Domain Adaptative Object Detection</b> (UDAOD) aims to alleviate the gap between the source domain and the target domain. Previous methods sought to plainly align global and local features across domains but adapted numerous pooled features and overlooked contextual information, which caused incorrect perceptions of foreground information. To tackle these problems, we propose <b>Diverse Feature-level Guidance Adjustments</b> (DFGAs) for two-stage object detection frameworks, including <b>Pixel-wise Multi-scale Alignment</b> (PMA) and <b>Adaptative Threshold Confidence Adjustment</b> (ATCA). Specifically, PMA adapts features within diverse hierarchical levels to capture sufficient contextual information. Through a customized PMA loss, features from different stages of a network facilitate information interaction across domains. Training with this loss function contributes to the generation of more domain-agnostic features. To better recognize foreground and background samples, ATCA employs adaptative thresholds to divide the foreground and background samples. This strategy flexibly instructs the classifier to perceive the significance of box candidates. Comprehensive experiments are conducted on Cityscapes, Foggy Cityscapes, KITTI, and Sim10k datasets to further demonstrate the superior performance of our method compared to the baseline method.
|
|---|