Method for estimating geometrical information of scene from single image through GAN (Generative Adversarial Network)

Abstract

The invention provides a method for estimating the geometrical information of a scene from a single image through a GAN (Generative Adversarial Network), and the method comprises the steps: inputtingan image of the scene and the depths of a plurality of pixels in the image into a trained GAN to obtain a depth image of the scene, wherein the depths of the pixels are the distances from the points,corresponding to the pixels in the image, in the scene to an observer, and the depth image is the sum of the depths of all pixels of one image. According to the invention, the method takes the image of the scene and the depths of a few of corresponding pixels in the image as the input, and employs a dual-consistency constraint GAN for predicting or estimating the depth image of the scene, so the method is simple and effective, and is low in cost.

Description

technical field [0001] The invention belongs to the field of computer image processing, and relates to a method for estimating scene geometric information from a single image, in particular to a method for estimating scene geometric information from a single image by using a generative confrontation network. Background technique [0002] Depth information prediction and estimation are very important in engineering applications, such as robotics, autonomous driving, augmented reality (AR) and 3D modeling. At present, there are mainly two methods for obtaining depth images, which are direct ranging and indirect ranging. Direct ranging refers to the use of various hardware devices to directly obtain depth information. For example, the TOF camera obtains the distance between the object in the target scene and the transmitter by emitting continuous near-infrared pulses; the laser radar scans the object in the measured scene by emitting laser light, and then obtains the distance ...

AI Technical Summary

Problems solved by technology

The invention adopts the structure of the fully convolutional network, removes the fully connected layer, and effectively reduces the parameter amount of the network. Although it can improve the problem of low resolution of the output image during the convolution process, the training samples required by this method are particularly large. long training time

Images