Lightweight Monocular Depth Estimation with Local Feature Enhancement Modules and Guided Data Augmentation

Although research on lightweight monocular depth estimation models has been actively conducted, achieving real-time deployment on edge devices remains challenging due to limited computational resources and memory capacity. To address these limitations, we propose a lightweight model for self-supervised monocular depth estimation. We cut the iterations of each feature-extracting block nearly in half, while our proposed Asymmetric Dilated Convolution module and a StarNext module compensate for the reduced model capacity. Specifically, the Asymmetric Dilated Convolution module captures horizontal and vertical structural features through asymmetric kernels, and the StarNext module fuses multi-scale features via element-wise multiplication. In model training, random cropping and scaling are applied for inducing the model to focus on localized object features. Additionally, we introduce a Disparity-guided Cutout technique based on the pre-inferred disparity map to randomly mask adjacent pixels. Simulation results on the KITTI dataset demonstrate that the proposed model reduces the number of parameters and GFLOPs by approximately 50% and 58%, respectively, without significant degradation in depth estimation accuracy compared to the baseline Lite-Mono. Furthermore, inference benchmarks on the Jetson Orin Nano platform demonstrate speedups of approximately 46.0%, 46.1%, 46.3%, and 52.0% across the MaxN, 25 W, 15 W, and 7 W power modes, respectively.