Domain generalization stereo matching method using monocular depth priors and feature consistency

Stereo vision simulates human binocular perception of scene depth and plays a critical role in applications such as autonomous driving, robotic navigation, and 3D reconstruction. Its key step, stereo matching, aims to establish pixel-wise correspondences between left and right views. To date, deep learning-based stereo matching methods have achieved excellent performance on standard datasets. However, due to the high cost of acquiring ground-truth disparity in real-world scenes, existing methods are often trained on synthetic datasets (e.g., SceneFlow). When these models are used in real-world scenarios, they often suffer from domain shift caused by factors such as differences in lighting conditions, variations in color and contrast, and scene types, leading to substantial performance degradation. To address this issue, this study investigates single-domain generalization in stereo matching.

This study proposes a domain-generalized stereo matching method based on depth prior guidance and feature consistency constraints (GCNet-FCDepth). The proposed method leverages the pretrained monocular depth model Depth Anything V2 to extract frozen multiscale geometric structural features, effectively capturing scene geometry without requiring additional stereo supervision. These geometric features are then fused with texture features via a selective channel fusion module, which adaptively selects the most informative channels from both feature streams to construct depth-aware matching features. This fusion strategy ensures that the resulting representation retains robust structural cues while remaining sensitive to surface details. To further enhance the matching process, a depth-guided dynamic cost aggregation module is designed. This module adaptively generates aggregation weights based on the affinity between disparity hypotheses and depth features. By incorporating geometric priors into the aggregation process, the module improves the reliability of the cost volume, especially in textureless regions and near depth discontinuities. During training, two complementary loss functions are introduced to explicitly constrain the domain invariance of matching features. The stereo contrastive loss encourages feature consistency between corresponding pixels across stereo views while suppressing irrelevant variations. The selective whitening loss promotes feature decorrelation and reduces domain-specific biases, thereby improving generalization across unseen environments.

Experimental results demonstrate that in zero-shot cross-domain generalization experiments, the error rate on the ETH3D dataset is reduced by approximately 64% relative to that of the current state-of-the-art real-time method RT-IGEV. On the KITTI 2015 benchmark, the proposed method achieves a D1-all score of 0.72 surpassing all compared methods in overall performance.

The proposed stereo matching method exhibits strong robustness in adverse weather scenarios such as fog and rain, thereby meeting the deployment requirements of real-world applications.