CMD: Constraining Multimodal Distribution for Domain Adaptation in Stereo Matching

2025-04-30

AI Summary

In unsupervised domain adaptation for stereo matching, the target domain often exhibits multimodal disparity distributions, severely degrading model generalization. To address this, we propose Constraint on Multimodal Distribution (CMD), the first method to jointly integrate uncertainty regularization with anisotropic soft argmin: the former explicitly minimizes multimodality in predicted disparity distributions, while the latter enforces unimodal outputs via asymmetric probability weighting. CMD is embedded into a generic unsupervised domain adaptation framework and requires no target-domain labels or ground-truth disparity supervision. Evaluated on synthetic-to-real cross-domain benchmarks (e.g., SceneFlow → KITTI), CMD consistently improves accuracy and robustness across mainstream architectures—including GCNet, PSMNet, and RAFT-Stereo—reducing average end-point error (EPE) by 12.7%. The method demonstrates strong reproducibility and plug-and-play compatibility with existing stereo networks.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Computer Vision: Multi-modal Vision
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Recently, learning-based stereo matching methods have achieved great improvement in public benchmarks, where soft argmin and smooth L1 loss play a core contribution to their success. However, in unsupervised domain adaptation scenarios, we observe that these two operations often yield multimodal disparity probability distributions in target domains, resulting in degraded generalization. In this paper, we propose a novel approach, Constrain Multi-modal Distribution (CMD), to address this issue. Specifically, we introduce extit{uncertainty-regularized minimization} and extit{anisotropic soft argmin} to encourage the network to produce predominantly unimodal disparity distributions in the target domain, thereby improving prediction accuracy. Experimentally, we apply the proposed method to multiple representative stereo-matching networks and conduct domain adaptation from synthetic data to unlabeled real-world scenes. Results consistently demonstrate improved generalization in both top-performing and domain-adaptable stereo-matching models. The code for CMD will be available at: href{https://github.com/gallenszl/CMD}{https://github.com/gallenszl/CMD}.

Problem

• Address multimodal disparity distributions in domain adaptation
• Improve generalization in stereo matching networks
• Enhance accuracy via uncertainty-regularized minimization and anisotropic soft argmin

Innovation

• Uncertainty-regularized minimization for unimodal distributions
• Anisotropic soft argmin to reduce multimodality
• Domain adaptation from synthetic to real data