Discriminative distribution alignment for domain adaptive object detection

Domain adaptive object detection has achieved appealing performance by constructing an effective transferable model for unlabeled target images, which takes advantage of the well-labeled source images with different distributions. However, two crucial factors are overlooked by most current methods: 1) different areas of an image should not be equally aligned since some areas may contribute more to distribution alignment if they contain more discriminative information for classifying the objects; and 2) the objectives of feature alignment and classification should not be independently optimized since it will fail to capture the discriminative information of data. To address these issues, we propose a new domain adaptive object detection model, referred to as discriminative distribution alignment domain adaptive detector. To be specific, the proposed method first makes the model focus on the areas that are quantified with high localization probability at the image level to enhance discrimination between foregrounds and backgrounds. Then the source and target images are aligned at the category level to learn class-invariant features by two adversarial regions-of-interest classifiers. Comprehensive experiments on several visual tasks verify that the proposed method outperforms the competitive domain adaptive object detection methods significantly in unsupervised domain adaptation setting. (c) 2021 Elsevier B.V. All rights reserved.