Diffusion-Driven Multiview Depth Estimation With Epistemic and Morphological Priors for Urban Scene Reconstruction

Abstract

Urban scene reconstruction plays a critical role in applications such as autonomous navigation and digital twin systems. As a key technique for scene reconstruction, multiview stereo (MVS) has made significant progress in general scenarios. However, it remains fragile in complex urban environments. Challenges such as weak textures, dense occlusion, depth adhesions, and illumination variations severely impair the effectiveness of traditional photometric and geometric consistency constraints, leading to unstable depth estimation and topological artifacts. To address these issues, we propose StableMVS, a robust depth estimation method that integrates semantic priors, structural cues, and perturbation-aware learning. Specifically, epistemic-structural prior infusion introduces cognitive priors and semantic layout from large vision models to enhance depth inference in ambiguous regions. Morphology-encoded depth field optimization (MeDF) leverages boundary-aware morphology to suppress distortions and topological artifacts near depth discontinuities. Furthermore, diffusion-propelled feature fidelity learning adopts a disturbance–restoration paradigm that improves feature stability under real-world perturbations such as lighting shifts and sensor noise. These components collectively form a unified framework that bridges global understanding with local consistency, yielding structurally coherent and resilient depth estimation across challenging urban scenarios. Experiments on multiple public datasets show that StableMVS achieves competitive accuracy under low-texture, occlusion, and illumination variation conditions, validating its effectiveness for real-world MVS applications. The code is available at https://github.com/IMOP-lab/StableMVS