Uncertainty-aware monocular-prior assisted stereo depth measurement for robust 3-D visual sensing

Abstract

Stereo vision is an important non-contact technique for 3D depth measurement, but its measurement reliability is often degraded in ambiguous regions, such as textureless surfaces, occlusion boundaries, and areas with weak visual correspondence. Although recent deep stereo methods have achieved high accuracy on standard benchmarks, their robustness under domain shifts and challenging measurement conditions remains limited. To improve the reliability of stereo depth measurement, this paper proposes MoGeStereo, an uncertainty-aware monocular-prior assisted stereo framework with binocular geometric arbitration. Specifically, monocular depth predictions from the pretrained vision foundation model MoGe-2 are introduced as prior cues to provide complementary structural information in regions where stereo correspondence is unreliable. Globally shared scale-and-shift alignment and pixel-wise uncertainty estimation are then used to convert the metric-scale monocular prediction into an aligned prior field. This prior field adaptively constrains the disparity search range and modulates the cost volume, allowing the matching process to focus on plausible disparities when the prior is reliable while reducing its influence in uncertain regions. A ConvGRU-based recurrent update module is further employed to refine the disparity estimate iteratively. During refinement, left–right consistency and photometric reprojection constraints are used for online geometric verification, ensuring that the final depth measurement remains governed by stereo geometry rather than by monocular prediction alone. Trained only on the synthetic scene flow dataset without additional real-world stereo supervision, MoGeStereo achieves strong zero-shot generalization on KITTI 2012/2015, ETH3D, and Middlebury. The results demonstrate that the proposed method improves cross-dataset transferability and enhances the robustness of stereo depth measurement in challenging visual sensing scenarios.