FSDE-net: Coarse-to-Fine Feature Separation Depth Estimation Network

Towards the drawbacks of insufficient accuracy and inexplicit details in monocular depth estimation, a coarse-to-fine feature separation depth estimation network with two stages, namely FSDE-net is proposed in this paper. A feature separation network to obtain one complementary group of features as salient features and depth features for distilling depth information is dedicatedly designed in the first stage. In the second stage, coarse depth maps of multiple receptive fields are generated for further restoring the refined depth map. In addition, a long-range skip connection structure between the two stages is proposed to compensate for information loss in the long propagation of the network. Experimental results on the KITTI dataset show that our proposed method can achieve the absolute relative error of 0.090, the squared relative error of 0.701 and the linear root mean squared error of 4.365, outperforming existing representative methods.