Method for compact 3D scene reconstruction based on semantic priori and progressive optimization of wide baseline

Abstract

A method for compact 3D scene reconstruction based on semantic priori and progressive optimization of wide baseline includes such step as providing multiple images with different visual angles, dividing all images into superpixel sets with local homogeneity and irregular shapes by superpixel segmentation, obtaining the initial depth of each superpixel in the image and the positional relationship between images from different angles of view, taking the high-level semantic a priori as a constraint condition, merging the adjacent superpixel regions in the same plane; using Markov random field model to estimate the depth of all the merged regions, obtaining the original depth map. By discarding the erroneous depth estimation and removing the redundant depth information from the original depthmap, the final three-dimensional scene can be obtained. The above methods can achieve more stable and accurate 3D reconstruction results than the traditional methods in different wide baseline environments.

Description

technical field [0001] The invention relates to the field of image processing, in particular to a wide-baseline dense three-dimensional scene reconstruction method based on semantic prior and progressive optimization. Background technique [0002] As a research hotspot in the field of computer vision, 3D scene reconstruction technology has been widely studied and applied in many fields such as aerospace, unmanned driving and digital entertainment. The traditional 3D scene reconstruction technology uses the motion-based structure recovery method (Structure From Motion, SFM) to recursively estimate the pose of the camera based on multiple image sequences taken from different perspectives and convert the scene into a sparse point cloud or a dense model. presented in three dimensions. One of the key issues in implementing this technique is how to accurately find the correspondence between images from different perspectives. Due to the randomness of the position and attitude of...

AI Technical Summary

Problems solved by technology

[0004] It is worth noting that triangulation-based geometry estimation methods during SFM require small camera motions between adjacent views, which is often not satisfied under wide-baseline conditions.

Images