Energy-based open set domain adaptation with dynamic weighted synergistic mechanism

Abstract

Open Set Domain Adaptation (OSDA) aims to minimize domain variation while distinguishing between known and unknown samples. However, existing OSDA methods, which rely on deep neural network classifiers, often lead to overconfident predictions and fail to clearly demarcate known from unknown samples. To address this limitation, we propose the Energy-based Open Set domain adaptation (EOS) method. EOS introduces a novel two-stage approach involving a separation stage followed by an alignment stage. In the separation stage, we employ an energy-based anomaly detection strategy to identify unknown samples, transforming the traditional K-class classification task into a K+1-dimensional classifier by introducing an additional dimension to model the uncertainty of out-of-distribution samples. To further refine separation, we apply a coarse-to-fine method that iteratively improves the separation outcomes, which are integrated as weighted inputs in the alignment process to enhance feature distribution alignment. In the alignment stage, we employ a dynamic weighted synergistic mechanism, where the separation network and alignment network co-evolve through continuous alternating training. This mechanism enables the system to better adapt to invariant features across domains. We evaluate EOS on standard benchmarks, including Office-31, Office-Home, and VisDA-2017, with experimental results demonstrating that EOS consistently outperforms other state-of-the-art methods.