Shape from inconsistent silhouette: Reconstruction of objects in the presence of segmentation and camera calibration error

Abstract

Silhouettes are useful features to reconstruct the object shape when the object is textureless or the shape classes of objects are unknown. In this dissertation, we explore the problem of reconstructing the shape of challenging objects from silhouettes under real-world conditions such as the presence of silhouette and camera calibration error. This problem is called the Shape from Inconsistent Silhouettes problem. A psuedo-Boolean cost function is formalized for this problem, which penalizes differences between the reconstruction images and the silhouette images, and the Shape from Inconsistent Silhouette problem is cast as a psuedo-Boolean minimization problem. We propose a memory and time efficient method to find a local minimum solution to the optimization problem, including heuristics that take into account the geometric nature of the problem. Our methods are demonstrated on a variety of challenging objects including humans and large, thin objects. We also compare our methods to the state-of-the-art by generating reconstructions of synthetic objects with induced error.

We also propose a method for correcting camera calibration error given silhouettes with segmentation error. Unlike other existing methods, our method allows camera calibration error to be corrected without camera placement constraints and allows for silhouette segmentation error. This is accomplished by a modified Iterative Closest Point algorithm which minimizes the difference between an initial reconstruction and the input silhouettes. We characterize the degree of error that can be corrected with synthetic datasets with increasing error, and demonstrate the ability of the camera calibration correction method in improving the reconstruction quality in several challenging real-world datasets.