Lightweight monocular depth estimation network for robotics using intercept block GhostNet

This study introduces a novel deep learning approach for monocular depth estimation that exhibits excellent performance while utilizing significantly fewer computational and memory resources. Our proposed method, called IBG-Mono, involves the use of the Intercept Block, in conjunction with cost-effective GhostNet components, to efficiently extract relevant information from the input image. The Intercept Block incorporates a mechanism that facilitates the retention and integration of low-resolution feature maps derived from the input image with the downsampled feature maps at different resolutions. On the other hand, the GhostNet components facilitate the efficient processing of coarse-grained feature maps by the Intercept Block, while also enabling their seamless integration with downsampled feature maps. Furthermore, a progressive downsampling is employed to maintain spatial alignment between feature maps of varying resolutions and those that have undergone downsampling. Our study involves conducting extensive experiments using NYU Depth V2 and KITTI dataset and presenting comparative results to the state-of-the-art lightweight monocular depth estimation with only 0.63M parameters and 0.31 GMACs. The results demonstrate the superiority of our system in comparison to other existing lightweight monocular depth estimation approaches.