LR-Depth: Enhancing Self-Supervised Monocular Depth Estimation in Low-Resolution Scenarios

Existing self-supervised monocular depth estimation methods have achieved impressive results by reducing reliance on expensive labeled data. However, their performance degrades significantly in low-resolution scenes due to challenges like missing edge details, difficulty in modeling blur regions, and inaccuracies in upsampling geometric structure reconstruction. To address this issue, we propose LR-Depth, an efficient self-supervised monocular depth estimation framework tailored for low-resolution scenes. Firstly, we propose a multi-scale wavelet edge enhancement that captures high-frequency edge information to improve depth estimation in edge regions. Meanwhile, to improve blurred regions, we leverage wavelet blur mapping to refine low-frequency component representation and enhance the modeling accuracy of depth distributions. Finally, we introduce geometric reconstruction upsampling that integrates position initialization with local offset constraints, enabling precise recovery of complex geometric structures while ensuring geometric consistency of the depth map. Extensive experiments on the KITTI dataset demonstrate that LR-Depth achieves state-of-the-art performance in low-resolution self-supervised monocular depth estimation, particularly excelling in edge clarity and depth prediction for blurred regions. Additionally, generalizability experiments on the Make3D dataset further confirm the framework’s robustness and broad applicability.