Self-supervised Monocular Depth Estimation Method Based on Piecewise Plane Model

Monocular depth estimation is crucial in scene understanding and autonomous driving. Recently, a considerable number of methods based on deep learning and piecewise plane geometric priors have made significant progress. However, these methods still face the following issues: 1) Current plane segmentation methods based on pixel color and object contours often yield sub-planes that are mostly curved surfaces, making direct fitting with a single plane model ineffective. 2) Plane prior knowledge ignores the depth distribution of pixels at sub-plane edges, leading to poor depth estimation at these edges.3) Existing methods often rely on difficult-to-obtain depth ground truth as supervision signals. To address the aforementioned issues, we propose a self-supervised monocular depth estimation method based on monocular video. We introduce a multi-plane fusion approach to fit sub-planes in images. We model the pixel depth at sub-plane edges as bimodal distribution and design a dynamic search method to enhance the computation efficiency of the cost volume. We validate the performance of our proposed method on the KITTI and NYU-Depth-v2 datasets.