Research On Graphs Based Geometric Model Fitting Methods

Model fitting is an important research subject in computer vision,and it has been applied in many areas,such as camera pose estimation,SLAM,image stitch,mo-tion segmentation,etc.The goal of model fitting is to estimate the number of model instances and the corresponding parameters in data.The observed data usu-ally includes a large number of outliers which puts forward high requirements for the model fitting methods.Recently,many researchers propose a variety of fitting methods with high robustness.However,current fitting methods are still not effec-tively in dealing with data seriously contaminated by outliers.Moreover,they face the problems of slow speed and low accuracy.Some researchers propose to remove outliers first.Then,they cluster the remaining data points and fit the corresponding models.However,current outlier removal methods achieve low accuracy.And they also have difficulties in dealing with data containing small structures.In response to the above problems,we propose to use residual information to obtain the relationships between data points and the relationships between model hypothe-ses.Then,we construct a simple graph model.This paper addresses issues such as simple graph const,ruction,the outlier removal methods,the model selection.Based on them,we propose some novel fitting methods.The main researching contents are summarized as following:(1)A weighted median-shift on graphs method is proposed.Current clustering-based fitting methods obtain the relationships between data points using preference information.Then,they cluster data points.However,these methods are not able to efficiently deal with the data points lying at the intersections,which may reduce the accuracy of fitting method.The weighted median-shift on graphs method obtains the relationships between model hypotheses using preferences information towards data points.And it weighs the model hypotheses via the kernel density estimation technique.Then,a simple graph is constructed.Finally,the proposed method clusters the model hypotheses by the weighted median-shift on graphs method.The proposed method efficiently combines the similarity measurements and weighting scores to cluster model hypotheses.Experimental results show that the weighted median-shift method is able to accurately estimate the number of models and the corresponding parameters in complex scene.(2)A dense subgraph detection based fitting method is proposed.Current density-based outlier removal methods treat the data points with low density values as out-liers,and it treats the data points with high density values as inliers.However,outliers will achieve high density values when outliers cluster together,causing high fitting errors.The dense subgraph detection based fitting method obtains the rela-tionships between data points by preference information towards model hypotheses,and it constructs a simple graph.Then,it maximizes the average similarities of point sets to guide the point sets shifting towards the density subgraph.The pro-posed method weighs data points by statistically analyzing the shifting results.And the data points with high weights are treated as inliers.Finally,Symmetric Nonnega-tive Matrix Factorization technique is applied to cluster inliers.Experimental results show that the dense subgraph detection based fitting method is able to effectively fit model instances from data containing small structures.