Category Dictionary Guided Unsupervised Domain Adaptation for Object Detection

Unsupervised domain adaption (UDA) is a promising solution to enhance the generalization ability of model from source target without manually annotating labels for data. Recent works in cross-domain object detection mostly resort adversarial feature adaptation match marginal distributions two domains. However, perfect alignment hard achieve and likely cause negative transfer due high complexity detection. In this paper, we propose category dictionary guided (CDG) UDA detection, which learns category-specific dictionaries represent candidate boxes domain. The representation residual can be used not only pseudo label assignment but also quality (e.g., IoU) estimation box. A weighted self-training paradigm then developed implicitly align domains training. Compared with decision boundary based classifiers such as softmax, proposed CDG scheme select more informative reliable pseudo-boxes. Experimental results on benchmark datasets show that significantly exceeds state-of-the-arts