CVGIP: Image Understanding最新文献

The problem of part definition, description, and decomposition is central to the shape recognition systems. We present a geometric model-driven framework for segmenting dense range data of complex 3D objects into their constituent parts in terms of surface (biquadrics) and volumetric (superquadrics) primitives, without a priori domain knowledge or stored models. Surface segmentation uses a novel local-to-global iterative regression approach of searching for the best piecewise biquadric description of the data. The region adjacency information, surface discontinuities, and global shape properties are extracted from biquadrics to guide the volumetric segmentation. Superquadric models are recovered by a global-to-local residual-driven procedure, which recursively segments the scene to derive the part-structure. A set of acceptance criteria provide the objective evaluation of intermediate descriptions and decide whether to terminate the procedure, or selectively refine the segmentation. The control module generates hypotheses about superquadric models at clusters of underestimated data and performs controlled extrapolation of part-models by shrinking the global model. Results are presented for real range images of varying complexity, including objects with occluding parts, and scenes where surface segmentation is not sufficient to guide the volumetric segmentation.

To date, the overwhelming use of motion in computational vision has been to recover the three-dimensional structure of the scene. We propose that there are other, more powerful, uses for motion. Toward this end, we define dynamic perceptual organization as an extension of the traditional (static) perceptual organization approach. Just as static perceptual organization groups coherent features in an image, dynamic perceptual organization groups coherent motions through an image sequence. Using dynamic perceptual organization, we propose a new paradigm for motion understanding and show why it can be done independently of the recovery of scene structure and scene motion. The paradigm starts with a spatiotemporal cube of image data and organizes the paths of points so that interactions between the paths, and perceptual motions such as common, relative, and cyclic are made explicit. The results of this can then be used for high-level motion recognition tasks.

This paper presents several sufficient conditions for a double or unique solution of the problem of motion and structure estimation of a rigid surface from pairs of monocular images. These conditions further the understanding of the uniqueness problem of rigid motion solution. We will show that five correspondences of noncolinear points that do not lie on a special type of quadratic curve, called a Maybank curve, in the image plane suffice to determine a pure rotation uniquely, and six correspondences of points that do not correspond to space points lying on a Maybank quadric suffice to determine a motion with nonzero translation uniquely. We will show that each Maybank quadric can sustain at most two physically acceptable motion solutions and surface interpretations, provided that a sufficient number of correspondences are present. In particular, we will show that in the plane motion case, six correspondences of points that do not lie on a quadratic curve in the image plane will admit only the true motion and structure and their duals as solutions. We will discuss how noise affects the uniqueness of solution and present a nonlinear algorithm for estimation of motion parameters. We will list several properties of the essential matrix T × R and the plane motion matrix R + TN^τ, both of which are frequently used in the motion and structure estimation problem. Simulation results are provided for verifying the theorems in this paper.

Model-based 3D object recognition systems using range imagery typically employ entirely data-driven procedures for segmentation and surface classification. However, some recognition environments may contain only objects whose surface types and parameters are known a priori and can therefore be exploited by the early-processing steps used in the recognition system. We propose a new suite of model-driven techniques for identification of quadric surfaces (cones, cylinders, and spheres) in segmented range imagery. The methods employ surface positions and surface normal estimates in combination with the known parameters of surfaces in a database of object models. Second-derivative quantities (i.e., surface curvatures) are not used. The free parameters of cylinders and spheres are accumulated using a Hough transform, and free parameters of cones are estimated using a regression procedure. Experiments are presented for numerous scenes of both real and synthetic objects including part jumbles, objects in many poses, objects containing concave and convex surfaces, and noiseless and noisy synthetic range images of objects. Our experimental results show that the proposed surface classification methods can accurately recover surface parameters from both synthetic and real images, making them viable for environments with partial knowledge of surface type and parameters.

The Hough transform is recognized as being a powerful tool in shape analysis which gives good results even in the presence of noise and occlusion. Major shortcomings of the technique are excessive storage requirements and computational complexity. Solutions to these problems form the bulk of contributions to the literature concerning the Hough transform. An excellent comprehensive review of available methods up to and partially including 1988 is given by Illingworth and Kittler (Comput. Vision Graphics Image Process. 44, 1988, 87-116). In the years following this survey much new literature has been published. The present work offers an update on state of the art Hough techniques. This includes comparative studies of existing techniques, new perspectives on the theory, very many novel algorithms, parallel implementations, and additions to the task-specific hardware. Care is taken to distinguish between research that aims to further basic understanding of the technique without necessarily being computationally realistic and research that may be applicable in an industrial context. A new trend in Hough transform work, that of the probabilistic Houghs, is identified and reviewed in some detail. Attempts to link the low level perceptive processing offered by the Hough transform to high level knowledge driven processing are also included, together with the many recent successful applications appearing in the literature.

We present two methods for detecting mirror symmetry in images, one based directly on the intensity values and another one based on a discrete representation of local orientation. A symmetry finder has been developed which uses the intensity-based method to search an image for compact regions which display some degree of mirror symmetry due to intensity similarities across a straight axis. In a different approach, we look at symmetry as a bilateral relationship between local orientations. A symmetry-enhancing edge detector is presented which indicates edges dependent on the orientations at two different image positions. SEED, as we call it, is a detector element implemented by a feedforward network that holds the symmetry conditions. We use SEED to find the contours of symmetric objects of which we know the axis of symmetry from the intensity-based symmetry finder. The methods described in this paper have been developed and tested for the recognition and tracking of cars in a real-time system for automatic car-following and headway control on normal roads.

This paper explores the possibility of using the binocular optic flow as an input for the correspondence process between stereoscopic images. The main advantage of the stereocorrespondence from optic flow (SCOF) is that it does not require the use of any a priori hypothesis concerning the 3D object under analysis. In order to determine its performance relative to noisy data, we applied an algorithm of SCOF on different rigid surfaces undertaking various 3D motions. We found that when SCOF is possible it is rather robust to noise. Moreover, the study of its domain of optimal efficiency shows that SCOF is likely to cooperate well with static stereopsis or structure from motion algorithms, thereby strengthening the processing of dynamic stereo images. As far as human vision is concerned, our psychophysical results indicate that a SCOF process does not seem to be used in the perception of 3D structure. This could be accounted for by the poor contribution of convergence signals to the perception of absolute depth in human vision, which seems incompatible with the precise knowledge of the geometry of the viewing system required by the SCOF.

Manipulation and transformation of symbolic images requires, in order to become efficient, various high-level data structures and procedures. This is especially true for applications of heuristic type. This paper introduces an image algebra which is based on symbolic projections and includes three fundamental aspects, i.e., a set of powerful relational operators, a generalized representation of "empty space" objects, and multilevel data representation for hierarchical data. Application to spatial reasoning is discussed.