Adaptive 3D Reconstruction Based On Joint Images

With the development of multimedia and computer,the technology of 3D reconstruction is now widely used in games,special effects of movies,digital city,digital museum,human-computer interaction,protection of non-material culture heritage,autonomous robot,medicine and other fields.Firstly,images of the object or the scene are obtained by cameras for 3D reconstruction.Then,according to the technologies of image processing,multiple view geometry and computer vision,the 2D information of images can be generated to 3D model of the object or scene.Nowadays,3D reconstruction based on vision is one of the most important technologies in the field of computer vision,which takes the most attention in this field.In order to get the 3D model from an interested image,more 2D information of images is expected to be obtained from the similar images,which are retrieved from the image set.Then the processing of 3D reconstruction is based on the image sequences that retrieving from the image set.In the above background,3D reconstruction based on joint images is researched,which combines the image retrieval and 3D reconstruction.For the reason of getting a perfect 3D model,image selection strategy is studied to select the appropriate images.The main contributions are listed as follows:(1)Firstly,the image retrieval is researched,which is based on the local SIFT feature.However,when extracting the SIFT feature,different number of local features are extracted for different images.For the efficiency comparison between different images,Bag of Words model based on SIFT feature are used to get better retrieval results.Besides,three kinds of similarity measurement that are Manhattan distance,Standardized Euclidean distance and Euclidean distance are discussed.What's more,precision and recall are used to evaluate the image retrieval system using the three kind of distance.(2)Multi-view 3D reconstruction and image selection strategy are studied in this section.There are information redundancies and errors in the image sequence from the image retrieval system.These information redundancies and errors lead to bad 3D models.Image selection strategy based on match points and homography matrix is studied,in choosing the most suitable images from the image sequence.Therefore,perfect 3D model can be obtained with small error.In the processing of reconstruction,which uses the method of Structure from Motion(SFM),the first step is to extract the SIFT features of the image sequence from the image retrieval system.The estimation of fundamental matrix,essential matrix and camera matrix are studied.However,the errors of the image matching points can result in serious error during computation of fundamental matrix,which can significantly affect the subsequent reconstruction.To solve this problem,RANSAC algorithm based on fundamental matrix is studied.RANSAC algorithm is utilized to filter the matching points and calculate the new fundamental matrix.To get the sparse cloud points,liner triangulation is used to calculate the coordinate values.But these sparse cloud points are invisible.Therefore,using PMVS dense match algorithm to make these cloud points visible.