Abstract
Panorama stitching has broken the limit of the eld of view in traditional image. In addition to watching from the web browser, Watching form VR (Virtual reality) headset make the distance between the panorama and user closer. However, the user can not move freely in the panorama and can only jump betwween the camera sets, which is limited by the positions of shooting cameras. Image-baesd rendering is a kind of technique for novel view synthesis that use the point cloud or disparty map as the aid to warp the source image to novel view and stitch them together. Their works can only handle the scenerios with small camera angle di erence or foucus on speci c kinds of objects. In this thesis, we adopt the similar idea of Chaurasia et al. [1] and Chang et al. [2] propose a system to synthesize any novel views. In order to design the algorithm of novel view synthesis directly, we simulate the source images and point cloud in virtual environment, oversegment the source image based on color in many superpixels, and merge these superpixel based on the object surface attributes. In the synthesis stage, we use a hierarchical warping strategy, which consists of homography wapring and non-linear warping, to warp the source images to the novel view. Finally these warped images are stitched together to render the novel view. We conducted an experiment to validate the e ectiveness of hierarchical warping and superpixel merging. The results show that the hierarchical warping can greatly reduce computation time and generates better results with merged superpixels.