HMV-Net: Resource-efficient monocular depth estimation for UAV cargo delivery

Monocular depth estimation for UAV cargo delivery must balance limited onboard computational resources against high accuracy requirements in safety-critical near-field regions (0-30 meters). Inspired by human myopic vision principles, this work presents HMV-Net, a bio-inspired framework that reallocates computational resources toward operationally critical foreground regions without introducing inference overhead. The framework implements two key components: a training-time attention mechanism that imposes no deployment cost, and a parameter-free refinement module for identified critical regions. Compared with uniform processing baselines, HMV-Net improves depth accuracy by 1.6%, reduces Root Mean Square Error (RMSE) by 2.1%, and enhances edge accuracy from 0.512 to 0.562, while maintaining identical parameter count (24.79M) and throughput (187 vs 192 FPS on desktop GPU). Embedded deployment on Jetson Orin NX achieves over 33 FPS with 1.8-5.0% accuracy improvements across NYU-v2, KITTI, and COCO2017, confirming practical effectiveness for resource-constrained UAV platforms. Code is available at https://github.com/tian1bukewei/HMV-Net.