Cross-Domain Object Detection with Hierarchical Multi-Scale Domain Adaptive YOLO

To alleviate the performance degradation caused by domain shift, domain adaptive object detection (DAOD) has achieved compelling success in recent years. DAOD aims to improve the model's detection performance on the target domain by reducing the distribution discrepancy between different domains. However, most existing methods are built on two-stage Faster RCNN, which is not suitable for real applications due to the detection efficiency. In this paper, we propose a novel Hierarchical Multi-scale Domain Adaptive (HMDA) method by integrating a simple but effective one-stage YOLO framework. HMDA-YOLO mainly consists of the hierarchical backbone adaptation and the multi-scale head adaptation. The former performs hierarchical adaptation based on the differences in representational information of features at different depths of the backbone network, which promotes comprehensive distribution alignment and suppresses the negative transfer. The latter makes full use of the rich discriminative information in the feature maps to be detected for multi-scale adaptation. Additionally, it can reduce local instance divergence and ensure the model's multi-scale detection capability. In this way, HMDA can improve the model's generalization ability while ensuring its discriminating capability. We empirically verify the effectiveness of our method on four cross-domain object detection scenarios, comprising different domain shifts. Experimental results and analyses demonstrate that HMDA-YOLO can achieve competitive performance with real-time detection efficiency.