Research On Image-based 3D Model Generation

This thesis mainly researched on the relative techniques of 3D modeling from the image sequence in multi-views. Image-based 3D modeling problem can be separate into two big problems: the first problem is how to segment object accurately, the second problem is how to reconstruct the 3D model of object from the image sequence in multi-view.Firstly, about object segmentation problem, the interactive object segmentation framework is proposed. On this framework, the user can select some pixels belong to object (foreground) and some pixels belong to background by the interactive tool, then the parameters of the model is learned or the energy function is optimized from these pixels to segment object more accurate. On this framework, we implement based-on machine learning object segmentation method and based-on graph cuts object segmentation method. Base-on machine learning object segmentation problem used SVM method and Boosting method, the kind of this method convert the segmentation problem to supervised learning problem, the parameters of the model is obtained by training from selected pixels, then use the model to predict the other pixels are belong to object or not, so the object is be segmented. Base-on graph cuts object segmentation method converts the segmentation problem to energy minimization problem, due to segmentation problem define the energy function, then construct the corresponding graph, find the min-cut (minimum energy correspondingly) of the graph, so the object is be segmented.Secondly, about 3D modeling problem, the stereo vision method and based-on silhouettes 3D modeling method are used. In this thesis, the stereo vision method use dynamic programming matching method to get the disparity of corresponding pixels, then use triangle calculation principle to get the depth information. In this thesis, a fast 3D model generation method from silhouettes is proposed, the traditional 3D cones intersection problem is converted to the 2D silhouettes intersection problem. Firstly, the 2D silhouettes of different viewpoints are projected back to parallel 3D planes, then the intersection silhouette of all back-projections on the 3D plane are calculated, finally corresponding points between the intersection silhouettes of two neighbor 3D planes are matched, so the mesh of 3D model are directly obtained. The 3D modeling from silhouettes algorithm mainly improves the accuracy of 3D model by increasing the number of viewpoints, moreover the time complexity of our algorithm is the linear time of the number of viewpoints. Theoretical analysis and experiment results both show that our method make out the 3D model effectively.