Geometric Model Fitting Algorithms Based On Consensus Analysis And Preference Analysis

Geometric model fitting is one of the most important research areas in computer vision.It is also a hot topic in the field of pattern recognition and image processing,which has been widely used in image registration,vanishing point detection and motion segmentation.The goal of geometric model fitting is to estimate the model instance that "best" explains the observed data.In fact,geometric model fitting is a very challenging problem in computer vision.In the past few decades,many domestic and foreign scholars have been committed to propose more robust and efficient model fitting methods.However,the current fitting methods have some problems in dealing with multi-structure data containing a large number of outliers,and they can not meet the requirements of practical applications in speed or accuracy.For the above difficulties,based on consensus analysis and preference analysis,two robust model fitting methods have been proposed to achieve good fitting results:(1)A conceptual space based model fitting method is proposed.Current preference analysis based model fitting methods use all the generated model hypotheses to represent data points.However,the generated model hypotheses often contain a large number of“bad”model hypotheses,which have a bad influence on the representation of data points.And this will make the computational complexity of those preference analysis based methods very high.The conceptual space based model fitting method constructs a conceptual space to describe the complex relationship between data points effectively by using only "good" model hypotheses.Based on the conceptual space,outliers can be well removed.In addition,the method also includes a model selection algorithm,which can quickly estimate the number and the parameters of model instances.The experimental results show that the proposed conceptual space based model fitting method can achieve better fitting results and high fitting efficiency in multi-structure data with large number of outliers.(2)A two-layer network based message passing fitting method is proposed.Consen-sus analysis based methods and preference analysis based methods have their own advantages and disadvantages.By combining the consensus analysis and preference analysis,the two-layer network based message passing fitting method construct a two-layer network.The network can describe the complex relationship between data points and model hypotheses.Based on the representation of the two-layer net-work,we futher proposes an effective two-stage message passing algorithm to fit the model instances in data.The experimental results show that the two-layer network based message passing fitting method can achieve better fitting results in complex multi-structure data.