Knowledge Graphs Meet Geometry for Semi-supervised Monocular Depth Estimation.

Depth estimation from a single image plays an important role in computer vision. Using semantic information for depth estimation becomes a research hotspot. The traditional neural network-based semantic method only divides the image according to the features, and cannot understand the deep background knowledge about the real world. In recent years, the knowledge graph is proposed and used for model semantic knowledge. In this paper, we enhance the traditional depth prediction method by analyzing the semantic information of the image through the knowledge graph. Background knowledge from the knowledge graph is used to enhance the results of semantic segmentation, and further improve the depth estimation results. We conducted experiments on the KITTI driving dataset, and the results showed that our method outperformed the previous unsupervised learning methods and supervised learning methods. The result of the Apollo dataset demonstrates that our method can perform in the common case.