Self-Supervised Monocular Depth Estimation Based on Full Scale Feature Fusion

In order to solve the problems of fuzzy boundary and artifacts in the depth map generated by self-supervised monocular depth estimation, a depth network coding and decoding structure based on full scale feature fusion module (FSFFM) and chain residual pooling module (CRPM) is proposed. For decoding in FSFFM, the higher resolution features, the same resolution features obtained by the encoder and the lower resolution features obtained by the previous decoder are fused with the upper level inverse depth map, which enables the features learned by the network contain both global and local information. Then CRPM is designed and used to get the background context information from the fusion features. Finally, an accurate depth map is obtained. Experiments are carried out on KITTI dataset. Compared with baseline, the absolute error of depth value is reduced by 7.8%, and the accuracy with a threshold of 1.25 is improved by 1.1%. The results are better than most existing self supervised monocular depth estimation algorithms.