3D Reconstruction of Multiple Objects from Dynamic Scenes and Learning Based Depth Super Resolution

Abstract

In this dissertation, a framework for reconstructing 3-dimensional shape of the multiple objects and the method for enhancing the resolution of 3-dimensional models, especially human face, are proposed. Conventional 3D reconstruction from multiple views is applicable to static scenes, in which the configuration of objects is fixed while the images are taken. In the proposed framework, the main goal is to reconstruct the 3D models of multiple objects in a more general setting where the configuration of the objects varies among views. This problem is solved by object-centered decomposition of the dynamic scenes using unsupervised co-recognition approach. Unlike conventional motion segmentation algorithms that require small motion assumption between consecutive views, co-recognition method provides reliable accurate correspondences of a same object among unordered and wide-baseline views. In order to segment each object region, the 3D sparse points obtained from the structure-from-motion are utilized. These points are relative reliable since both their geometric relation and photometric consistency are considered simultaneously to generate these 3D sparse points. The sparse points serve as automatic seed points for a seeded-segmentation algorithm, which makes the interactive segmentation work in non-interactive way. Experiments on various real challenging image sequences demonstrate the effectiveness of the proposed approach, especially in the presence of abrupt independent motions of objects. Obtaining high-density 3D model is also an important issue. Since the multi-view images used to reconstruct 3D model or the 3D imaging hardware such as the time-of-flight cameras or the laser scanners have their own natural upper limit of resolution, super-resolution method is required to increase the resolution of 3D data. This dissertation presents an algorithm to super-resolve the single human face model represented in 3D point cloud. The point cloud data is considered as an object-centered 3D data representation compared to the camera-centered depth images. While many researches are done for the super-resolution of intensity images and there exist some prior works on the depth image data, this is the first attempt to super-resolve the single set of 3D point cloud data without additional intensity or depth image observation of the object. This problem is solved by querying the previously learned database which contains corresponding high resolution 3D data associated with the low resolution data. The Markov Random Field(MRF) model is constructed on the 3D points, and the proper energy function is formulated as a multi-class labeling problem on the MRF. Experimental results show that the proposed method solves the super-resolution problem with high accuracy.