Open set domain adaptation via unknown construction and dynamic threshold estimation

Abstract

Open set domain adaptation (OSDA) focuses on adapting a model from the source domain to the target domain when their class distributions differ. The goal is to accurately recognize unknown classes while correctly classifying known classes. Existing research has indicated that adversarial networks can be efficient for unknown class recognition, yet threshold setting remains a challenge. We address this challenge by proposing an OSDA method that uses unknown construction and dynamic threshold estimation (UCDTE), which consists of three stages: unknown construction, dynamic threshold estimation, and distribution alignment. In the first stage, known as unknown construction, pseudo-unknown samples are constructed through feature fusion to learn information regarding the unknown class. In the second stage, dynamic threshold estimation, an unknown discriminator is constructed to further explore different semantic information in the unknown classes, and a dynamic threshold is generated for each target sample by combining it with the domain discriminator. Finally, in the distribution alignment stage, the dynamic threshold adversarial network aligns known samples between the source and target domains while reducing the intra-class gap of unknown samples in the target domain. Experiments conducted on three datasets have demonstrated the robustness and effectiveness of our approach in adapting models across different domains.