Mapping Method between 2D Landscape Image and 3D Spatial Data based on Adversarial Relative Depth Constraint Network

Abstract

This paper proposes a mapping method of 2D image and 3D spatial data based on the adversarial relative depth constraint network. The steps are as follows: 1) Input pixel coordinates of key nodes of 2D landscape image, and conduct normalization preprocessing; 2) Input two-dimensional pixel coordinates into the depth prediction network and output the depth values of key nodes; 3) Using depth values and two-dimensional pixel coordinates to reconstruct three-dimensional coordinates of key nodes; 4) Input DEM data to the discriminator of the generated adversarial network to calculate the authenticity error value, and use the relative depth information between the attitude characteristics of mountain and hydrology and the corresponding key nodes of the image to calculate the relative depth error; 5) Add the authenticity error and relative depth error calculated above to get the total error, and feed back to the depth prediction network to get a more accurate mapping evaluation, so as to realize mapping discovery. The problems solved in this paper include: lack of characteristic pose data in the traditional geo-evidence-based process of 2D landscape images; The results of the generative adversarial network method do not conform to the relative depth relationship of feature points in 3D spatial data.