GECL: Fine-grained granular envelope contrastive learning for unsupervised domain adaptation.

Abstract

Unsupervised domain adaptation (UDA) seeks to transfer knowledge from labeled source data to an unlabeled target domain under distribution shifts. Existing class-aware UDA approaches alleviate negative transfer by leveraging label information. However, they often overlook intra-class diversity and concept shift, limiting their ability to capture fine-grained semantic structures. In this work, we propose a novel Granular Envelope Contrastive Learning (GECL) method that explicitly models intra-class variations by generating multiple granular envelopes for each class. Firstly, a granular envelope generation mechanism is introduced that recursively partitions the feature space based on a local purity criterion. These envelopes act as refined class prototypes, enabling more accurate characterization of class distributions. Secondly, a sample-to-envelope contrastive learning objective is designed to enhance discriminative feature representation. Thirdly, an envelope-guided consistency regularization strategy is employed to enhance the semantic consistency between model predictions and envelope structures. By incorporating these components into an envelope-aware optimization framework, the proposed method jointly reduces domain discrepancy and enhances feature discriminability. Extensive experiments on multiple benchmarks show that our approach achieves state-of-the-art performance across diverse domain shift scenarios, especially under large-scale or high-divergence settings.