A Review of Techniques for Lightweight Monocular Depth Estimation

摘要

Recent advancements in lightweight neural networks have enabled the deployment of monocular depth estimation (MDE) models on embedded platforms, including microcontrollers, mobile devices, and edge computing environments. Importantly, it allows the effective implementation of various applications, including visual synchronous localization and mapping (SLAM) in resource-constrained robotic systems. This review examines multiple approaches currently being explored for lightweight monocular depth estimation models, with a focus on techniques such as architecture optimization, network pruning and slimming, hybrid convolutional neural network (CNN)-transformer architectures, knowledge distillation, quantization, and low-precision computation, efficient training strategies, innovative loss functions, edge information utilization, energy-quality scalability, and improvements in up-sampling and decoders. This review highlights that combining architecture optimization, network pruning, and slimming with quantization and low-precision computation provides the optimal approach for reducing model parameters while preserving high accuracy.