Improving Localization Accuracy For Domain Adaptive Object Detection

Object detection has achieved great success with the development of deep learning techniques and a large number of training data.However,as there exist large variances in object appearance,background environments,light illumination and image quality between the benchmark datasets and the real-world images,the performance of the object detectors trained on the benchmark datasets(source domain)degrade dramatically when detecting real-world images(target domain).This phenomenon is considered as a domain shift problem.The most direct approach to solve domain shift is to acquire annotations for target domain.However,it consumes a large amount of time and effort to annotate datasets.Therefore domain adaptation is proposed to solve it without creating annotations for target domain,which can be categorized into four types as instance-based,mapping-based,network-based and adversarialbased.However,as these methods are originally designed for classification task,they are not fully competitive with object detection task,which requires more accurate localization of the objects.Detection methods that directly adopt adaptation approaches for classification will suffer from the degradation of the adaptation and inaccurate object localization.In this paper,we present a novel deep neural network design for domain adaptive object detection by further improving the localization accuracy of objects.First,we insert several residual blocks into the shallow layers of a convolutional neural network used in the target domain to enhance detailed spatial information,which helps for object localization.Second,we present to refine the pseudo labels generated from the current object detection methods and use these labels with a weighted loss function to train the network on target domain.We perform various experiments to evaluate our network on three widely-used public benchmark datasets for domain adaptive object detection.Experimental results show that our proposed method performs favorably against state-of-the-art methods on all the datasets quantitatively and qualitatively.