Confused and disentangled distribution alignment for unsupervised universal adaptive object detection

Abstract

Universal domain adaptive object detection (UniDAOD) is a more challenging and realistic problem than traditional domain adaptive object detection (DAOD), aiming to transfer the knowledge from the well-labeled source domain to the unlabeled target domain without any prior knowledge of label sets. Intuitively, the main challenge of UniDAOD is to eliminate the domain shift and suppress the interference caused by the category shift induced by private classes (i.e., classes only existed in one domain). In the current study, we propose a simple but effective CODE framework, namely Confused and Disentangled Extraction, for alleviating this issue. Specifically, we propose the virtual adversarial adaptation module, characterized by incorporating virtual domain labels within the domain classifier for unaligned samples. This confuses the domain classifier, effectively addressing the issue of converging to local optima resulting from equilibrium challenges and consequently narrowing the domain shift. Simultaneously, we introduce the entropy margin separation module, which utilizes the distinctiveness of category predictions as a disentangled factor. This enables the automatic discovery of private classes in each domain, suppressing interference during the adaptation process. Experiments on four universal scenarios (i.e., closed-set, partial-set, open-partial-set, and open-set) show that CODE obtains a significant performance gain over original DAOD detectors.