Unsupervised cross-domain object detection based on dynamic smooth cross entropy
摘要
Pseudo-label self-training methods are highly regarded in cross-domain object detection (CDOD) models for their reduced annotation costs, strong applicability, and straightforward deployment. However, existing methods still struggle to balance the quantity and quality of pseudo-labels, as well as their efficient utilization. To tackle these challenges, this study introduces a novel approach by integrating strong-weak data augmentation (SWDA) techniques into the decoupled adaptation framework, significantly enhancing the generalization ability and accuracy of the pseudo-label generator. Furthermore, inspired by Softmatch, this research proposes a dynamic smooth cross-entropy loss function that adjusts pseudo-label confidence thresholds in real-time based on the model’s learning state and the predictive distribution across categories. This method not only retains low-confidence pseudo-labels but also assigns varied weights to them according to their confidence levels, enhancing both the quality and utility of pseudo-labels. This approach substantially improves CDOD performance, especially demonstrated in the Pascal VOC