Image Segmentation In Multiple Views And3D Reconstruction

With the development of feature point detection and image segmen-tation, the3D reconstruction technology is becoming easier and easier tobe implemented. If everyone can build3D models of surrounding objectseasily, user experience in areas such as online shopping, social networkcan be greatly improved. In this thesis, we implemented a realtime3D re-construction system, which can generate models with high quality throughuser interaction.To achieve this, we have three main steps. Firstly, we implemented ahigh efciency image segmentation tool. By optimizing the Gibbs func-tion, the user can grab the object from the background using just a littleinteraction. To make the segmentation process run faster, we adopt the s-trategy which frst downsamples the initial image and then upsamples it.Thisstrategycanhelpreducethereactiontimegreatlywhentheimagesizeis large. Secondly, we implemented a viewing cone based3D reconstruc-tion algorithm. We intersect the viewing cones derived from the imageswhich are taken by the user under various angles. In this way, we can getthe bounding box of the object. After that, we map texture onto the bound-ing box and get the fnal model. We also transplanted the algorithm toGPU, so the whole process can be run in realtime. The execution speedof GPU version is10times faster than the one of CPU version. Finally, to improve user experience, we record the information of user’s editing andupdate all the images in background automatically based on the historyinformation. In such a way, the human labor is released.In our experiments, we fnd that our system can run in high efciencywhile build3D model with less user interaction.