Semantics Driven 3D Shape Analysis And Modeling

The rapid development of digital geometry processing technique has been greatly boosting the growth of digital geometry. As a newly emerging digital multimedia, geometry is populating the internet with a remarkably speed. The significant growing of digital geometry in terms of both quantity and quality calls for more effective ways of processing and utilization. To this end, geometry processing is currently moving towards high-level shape analysis and understanding, aiming at discovering the underlying semantic information of a 3D shape. This gives the birth of the recent trend of high-level geometry processing, which has been widely recognized by the graphics community.Shape semantics reflects human knowledge about shape's geometry, structure, functionality, and their relationship. The goal of shape analysis is to automatically extract such knowledge from a shape with the help of both geometry processing and input knowledge. Effective integration of the input shape knowledge is important to shape analysis. We first study the symmetry analysis of a single shape and develop symmetry-driven mesh segmentation for structural analysis. We then study the consistent analysis of semantics of a 3D model set. The pre-analyzed semantics greatly facilitate 3D modeling driven by a model set. The main contributions of this thesis include:1. Detection of partial intrinsic reflectional symmetry. Symmetry bridges the gap between low level geometry and high level semantics. Therefore, symmetry analysis is one of the central problems of shape analysis. Previous works have studied the detection of global extrinsic, partial extrinsic, and global intrinsic symmetries. Partial intrinsic symmetry is more general in 3D shapes, although its detection is more difficult due to the complexity of its parametric representation. This thesis, for the first time, studies the problem of automatic detection of partial intrinsic reflectional symmetry (PIRS) on a closed 2-manifold. We begin with a formal definition of PIRS on a closed 2-manifold using the maximal generating set of intrinsic reflectional symmetry. Based on the definition, we propose a robust, voting-based detection algorithm for PIRS over a 3D triangle mesh. To explicitly extract a set of intrinsic reflectional symmetry axis (IRSA) curves, we propose an iterative grass-fire region growing method. With the IRSA curves and their corresponding regions of generating set, we achieve symmetry-aware segmentation.2. Co-analysis of a model set. A set of models belonging to the same class often contains more semantic information than a single one; they often share the same functional parts, which may help better understanding the structure of the shape class. We study the problem of co-analysis of a set of man-made objects belonging to a certain class and present a framework for shape style analysis. Shape style is defined based on the anisotropic part scales. We perform an unsupervised style clustering. Through intra-style co-segmentation and inter-style part correspondence, we achieve style-content separation for the input shapes, where content refers to part composition and geometry and style represents scale proportion between functional parts. With such separation, we arrive at a consistent segmentation of the input set.3. Shape synthesis based on style transfer. With the consistent segmentation of a set of 3D shapes, we propose shape synthesis by style transfer between any two shapes in the set. Style transfer is achieved by part scaling according to the part correspondence implied by consistent segmentation. However, part scaling may detach neighboring parts. To re-align and stitch the detached neighboring parts, we propose an iterative re-alignment approach based joining point set between the two parts, as well as a joining curve aware method for natural part stitching.4. Photo-inspired data-driven shape modeling. Co-analysis augments a 3D model set with rich semantics. With the help of this, we introduce an algorithm for 3D object modeling where the modeling inspiration is drawn from an object captured in a single photograph. The modeling process is supported by an available set of candidate models, which have been pre-analyzed to possess useful high-level structural information. Our method creates a digital 3D model as a geometric variation from the candidate that best resembles target object in the photograph. To facilitate modeling from image, we propose image-space object segmentation, candidate model retrieval using labeled segmentation of the input image, together with a silhouette driven structure-preserving shape deformation. The main features of our method are three folds. First, the whole modeling pipeline makes heavy use of the semantics pre-analyzed for the candidate set, which helps to compensate for the ill-posedness of the 2D-to-3D reconstruction from a single image. Second, the structural information is preserved by the geometric variation so that the final product is coherent with its inherited structural information readily usable for subsequent model refinement. Third, also due to structure preservation, the resulting 3D model, although built from a single view, is structurally coherent from all views.