Designing and Searching for Lightweight Monocular Depth Network

Depth sensing is extremely notable in some tasks of robot and autonomous driving. Nowadays, monocular depth estimation based on deep learning becomes a research focus in computer vision. However, most of the current work is seeking for more complex models to get higher accuracy, which can not achieve real-time inference on mobile or embedded systems. Therefore, we aim to design a lightweight model in this paper. At first, we improve the state-of-the-art model, FastDepth, producing FastDepthV2, which has higher accuracy and lower latency on the NYU Depth v2 dataset. Besides, since designing artificial networks takes time and effort, we make it automatic to design lightweight models in monocular depth estimation, using neural architecture search (NAS). Further, inspired by the architecture of MobileNetV2, a factorized hierarchical search space for encoder is used in this paper. In the meanwhile, we incorporate the accuracy and multiply-add operations of a model together into the searching objective, and use gradient-based reinforcement learning algorithm in the searching iterations. The controller in the reinforcement learning framework has converged after more than 1000 searching iterations, and three network architectures with the best performance are obtained. Under the same conditions of training and testing, two of them perform better than FastDepthV2.