MAD: Mask-Guided and Dual-Model Self-Training for Unsupervised Domain Adaptation in Stereo Matching

Abstract Stereo matching models trained on synthetic data often degrade when transferred to real scenes due to the domain gap. Unsupervised domain adaptation provides a practical way to exploit unlabeled target-domain stereo images without requiring additional disparity annotations. However, existing adaptation approaches often introduce architecture-specific components or additional learnable modules, which can limit their applicability across different stereo matching networks. To address this issue, we present MAD, a simple, architecture-agnostic adaptation framework that combines mask-guided self-supervised fine-tuning with dual-model pseudo-label self-training. In the first stage, left-right consistency of the source-domain pretrained model is used to identify reliable non-occluded regions and reduce the influence of unreliable photometric supervision during target-domain fine-tuning. In the second stage, the predictions of the pretrained and fine-tuned models are jointly considered, and geometric and photometric consistency checks are applied to select filtered pseudo-labels for further self-training. The framework does not require modifications to the internal architecture of the underlying stereo network or additional trainable adaptation modules. We evaluate MAD on multiple real-world datasets and across several stereo matching backbones, including PSMNet, GwcNet, ACVNet, CoEx, and CFNet. The results show consistent improvements over conventional unsupervised fine-tuning across a range of target domains and environmental conditions. Cross-domain experiments further indicate that the adapted models can retain useful transferability across different target domains. Overall, combining simple geometric filtering with complementary predictions provides a practical, architecture-independent strategy for synthetic-to-real stereo adaptation.