Neural network training method and system for three-dimensional reconstruction

Abstract

The invention provides a neural network training method and system for three-dimensional reconstruction, and the method comprises the steps: carrying out unsupervised calibration of inputted left and right images of a camera, obtaining a plurality of disparity maps through continuously reducing the left and right images of the camera, and respectively calculating the variance of corresponding points of the disparity maps; and acquiring a rough disparity map to train the deep convolutional neural network, so that the latest convolutional neural network DCNN for three-dimensional reconstruction can easily adapt to new conditions in practice, and a large amount of manual calibration is avoided.

Description

technical field [0001] The present invention relates to the field of machine learning, in particular to a neural network training method and system for three-dimensional reconstruction. Background technique [0002] Human eyes take pictures separately and notice the difference that gives us the perception of depth. This difference is called parallax. The parallax is inversely proportional to the depth. Once the parallax information is obtained, the depth information of the object can be calculated through the formula, and the two-dimensional picture becomes three-dimensional. The two calibrated cameras obtain the left image of the camera and the right image of the camera respectively, and the horizontal difference between the corresponding points of the two images is the parallax. A correctly labeled map (ground truth) is required to determine the corresponding point of each point on another map. In order to obtain the correct ground truth, it is generally necessary to man...

AI Technical Summary

Problems solved by technology

In order to obtain the correct ground truth, it is generally necessary to manually calibrate a large number of points, which is time-consuming and laborious

[0003] In recent years, most technologies based on computer vision and deep learning adopt passive methods, such as inputting single or multiple images from a single viewpoint, and deriving scene depth information through two-dimensional features such as light and shade, shadows, etc., but Average reconstruction effect

Others use monocular images marked with detection data to train the target detection neural network. Since the monocular image does not have accurate three-dimensional depth information and scale information, the accuracy of the trained target detection neural network is low; or The depth information of the monocular image can be determined by means of lidar, depth model, etc., and then the depth information and the monocular image marked with detection data can be used to train the target detection neural network. Because the acquisition of depth information is more complicated, the detection of the target object is difficult. The process is also more complicated, the detection time is longer, and the detection efficiency is low

Images