Mahalanobis distance-guided conditional adversarial learning for universal domain adaptation

Cangning Fan, Peng Liu, Wei Zhao

Abstract: Highlights•A novel UniDA framework is proposed based on Mahalanobis distance and conditional adversarial learning.•Mahalanobis-based scoring enables effective private class detection through density estimation and generative modeling.•A Mahalanobis-guided conditional adversarial method is designed for multimodal and discriminative feature alignment.•Extensive experiments on UniDA benchmarks demonstrate superior performance and validate the proposed components.