EdgeMono: A Lightweight Self-Supervised Monocular Depth Estimation Framework for Edge Devices

Self-supervised monocular depth estimation is a promising solution for edge devices of the Internet of Things due to its label-free training. However, current methods are often hampered by two critical challenges: degradation of image quality and loss of high-frequency details. To address these issues, we propose EdgeMono, a lightweight architecture that integrates a Degradation-related Prior Prompt Generation (DPPG) module to rectify features impacted by low-quality inputs, along with a Multi-scale Mixture of Experts Network (MMEN) and a Frequency-assisted Progressive Feature Enhancement Module (FPFEM) to recover fine-grained depth details. Experiments on the KITTI dataset demonstrate that EdgeMono reduces the AbsRel error by 23.6% and the model size by 79.5% compared to Monodepth2, confirming its efficacy for real-time depth estimation on resource-constrained devices.