Image Multi-plane Detection Based On Region Matching

There are many planar structures in our daily life,and these structures play an increasingly important role in computer vision relying on its special structural characteristics.It is widely used in image registration,camera calibration,3D reconstruction,augmented reality and so on.One of the most common methods of plane detection is the motion feature of pixel pairs between different visual angles.Thus,we can use homography to define a plane model,and estimate multi-models according to the association between feature point pairs and plane models.These traditional methods only use the coordinate offset of the feature point pairs to calculate the matching score,which limits the accuracy and robustness of the plane detection.On the other hand,we can only get the classification of discrete feature points instead of continuous pixels for the whole image.At the same time,smooth plane without texture can not find enough feature points,so that this plane can not be extracted.Moreover,fixed threshold of plane detection can not satisfy multiscale planar division from different scenes.In order to solve these problems,we present a new plane detection method based on region matching.Firstly,we take advantage of line and region matching to optimize RANSAC,which can improve the accuracy and precision of model fitting,and generate more effective plane model candidates.Next,we allocate the best matched plane candidate for each initial region by region matching method.According to the coplanarity of the different regions,we merge adjacent regions hierarchically,and construct a hierarchical tree of plane extraction.Finally,we can flexibly choose any level to get plane division of whole image.In this paper,we try to combine region information with model fitting,and transform traditional plane model fitting problem into hierarchical clustering of region merging problem.Experimental results prove the good performance in plane detection task of complex scenes.