MBUDepthNet: Real-Time Unsupervised Monocular Depth Estimation Method for Outdoor Scenes

Abstract

Monocular depth estimation technology has emerged as a critical component across a variety of outdoor applications like robotics, augmented reality, autonomous driving, and 3D reconstruction. Mainstream monocular depth estimation methods consistently face challenges in applications requiring real-time performances, as they exhibit considerable computational complexity, resulting in poor runtime performance. Here, we propose an innovative processing module named MDE-Lite. Based on that, we develop a lightweight yet effective depth estimation network named MBUDepthNet. Besides, we build a training scheme with multiple loss functions. Experimental validation on KITTI dataset demonstrates that our method not only rivals mainstream methods in terms of accuracy but also exhibits superior computational efficiency. Compared to the method using ResNet-18, our method achieves a 22% higher frame rate in terms of frames per second.