Optimizing Monocular Depth Estimation for Real-Time Edge Computing Platforms

This paper presents an efficient approach for low-latency monocular depth estimation on edge systems with low power consumption. Leveraging insights from human depth perception, we propose EfficientDepth-Small, a novel network architecture optimized for real-time operations. Our design achieves a remarkable balance between accuracy and runtime, utilizing a lightweight encoder-decoder architecture, depthwise decomposition, and pruning techniques to reduce model complexity without compromising competitive accuracy. Hardware-specific compilation further enhances runtime performance, resulting in significant reductions in inference latency on the NVIDIA Jetson Nano platform. Experimental results demonstrate superior accuracy and runtime efficiency compared to previous studies, with our approach achieving a δ1 (Depth Accuracy) score of 90.1% and a runtime of 23.40 ms on the Jetson Nano GPU. Power consumption measurements reveal successful real-time depth inference at under 10 watts of active power. Overall, our study advances the state-of-the-art in monocular depth estimation and opens avenues for future research in edge computing and computer vision applications.