Monocular depth estimation system implementation method based on adversarial network

Abstract

The invention discloses a monocular depth estimation system implementation method based on an adversarial network, and relates to monocular image depth estimation. The method comprises a training stage and a test, and the specific steps of the training stage are as follows: collecting an RGB image set for training and a depth image set corresponding to the first n images; initializing a generatornetwork parameter theta G and parameters theta PD and theta DD of the two discriminator networks according to a designed network structure; setting the number of iterations of training. The specific steps of the test stage are as follows: importing a trained generator network weight; transmitting one image as input to a generator network; and calculating the input image by using the imported network weight, and outputting depth value information corresponding to each pixel point in the input image through the network. The bottleneck of active depth perception hardware depth holes and depth sparseness is broken through, high depth recovery precision is kept, and the method has important practical value and significance in the fields of scene reconstruction, unmanned driving, augmented reality and the like.

Description

technical field [0001] The present invention relates to monocular image depth estimation, in particular to an implementation method of a monocular depth estimation system based on an adversarial network. Background technique [0002] When humans see an image, they can often easily understand the three-dimensional structure of the scene. However, inferring the 3D structure of a scene remains a difficult problem for current computer vision systems. Indeed, from a narrow mathematical perspective, it is impossible to recover the 3D structure from a single image, since we have no way of knowing that an image was taken against the photograph (in which case all depths should be in on a plane) or against a real 3D scene. But in real life, people can have a very good perception of the depth of the scene in the photo just through a photo, and among all possible depth values, most of the possibilities are impossible in the real world. Therefore, depth in monocular images can still b...

AI Technical Summary

Problems solved by technology

However, inferring the 3D structure of a scene remains a difficult problem for current computer vision systems

Indeed, from a narrow mathematical point of view, it is impossible to recover the 3D structure from a single image, because we cannot know that an image was taken against the photo (in which case all depths should be in on a plane) or against a real 3D scene

Images