Research on edge-aware lightweight monocular depth estimation methods

• 1、College of Computer Science (National Pilot Software Engineering), Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：This paper proposes an edge-aware lightweight monocular depth estimation method to alleviate the deployment bottleneck of high-precision foundation depth models on resource-constrained edge platforms, where large parameter scale and high computational cost hinder real-time inference. The goal is to achieve dense depth prediction with both strong cross-scene generalization and low inference latency. To address boundary blurring caused by attenuation of high-frequency spatial details during channel-capacity compression in lightweight models, improvements are made at both feature modeling and supervision levels. First, an edge-attention mechanism is designed to adaptively enhance responses related to geometric boundaries in high-frequency features. Second, a multi-cue edge supervision scheme is constructed by integrating depth discontinuity and surface-normal gradient information, thereby improving depth sharpness and geometric completeness at object contours and fine structures. Experimental results show that the proposed model achieves a $\delta_1$ accuracy of 0.964 in zero-shot evaluation on NYU Depth v2, and reaches real-time inference speeds of 65/15 FPS on NVIDIA Jetson Orin NX/Nano edge devices, respectively. By combining knowledge distillation from a high-precision model with edge-aware modeling, this study provides a feasible path for engineering deployment of monocular depth estimation in resource-limited scenarios.

Keywords： monocular depth estimation lightweight model edge awareness knowledge distillation