Unsupervised single image-based depth estimation powered by coplanarity-driven disparity derivation

Nowadays, depth estimation has become a mature technique and achieved tremendous progress. However, we observed that existing depth estimation methods perform poorly in reconstructing local three-dimensional structures with rich repeated textural and structural patterns, primarily due to insufficient modeling of planes, which leads to significant depth estimation errors at non-coplanar edges. In this work, we propose an unsupervised depth estimation method using the coplanarity for disparity inference. First, a local coplanaritydriven disparity derivation module is conducted to learn the plane parameters for each pixel, providing the network with an initial understanding of the scene's planar structure. Second, a disparity map is generated using loss constraint network, including reconstruction matching, disparity smoothness, and left-right disparity consistency. Third, an unsupervised architecture based on binocular image pairs is constructed to remove any potential adverse effects due to unknown scale or estimated pose errors. Forth, with the known baseline distance and camera focal length, the disparity map is converted into the depth map to perform the end- to-end depth estimation from a single image. In the end, extensive experiments on both ancient Chinese architectures and benchmark datasets demonstrate the high accuracy and robustness of our method in depth estimation, confirming its practical applicability in real-world engineering. Source code is available at https: //github.com/yaozz1110/C3DSIDE.