The robust detection of curvilinear structures from contours represents an important cue for most computer vision systems. However, the analysis of 2D images of a scene is a difficult task, made of under constrained, ill posed problems. Perceptual Organization is one contribution of psychovisual theories to computer vision. It provides generic approaches for the perception of scenes, necessary to reduce constraints and solve ambiguities.
The purpose of this paper is to propose an efficient method, based on Gestalt rules, to detect salient contours and extract elements of representation from complexe scenes. The proposed grouping strategy is hierarchical and divided in three levels of organization.
The first level is based on the saliency network framework. A quality function based on curvature, co-circularity, continuity and orientation is optimized throughout a network of locally connected elements. The purpose of this level is to reduce the complexity of visual tasks by selecting the most salient linear structures from contours. We propose a generic formalism for the conception of such networks, a more stable family of quality functions, a new algorithm for optimization and criteria for selection of a set of salient groups after optimization. Salient groups of edge elements play a role of focus of attention for the second level of organization. Hypotheses of segments, arcs and points of interest are proposed from each salient group and grouped following rules of parallelism, proximity, continuity to create a reduced set of representative elements of contours. Finaly, we illustrate the final level of organization with an application to junction detection and matching on digital and real scenes.
The main caracteristic of our approach is the separation between a general strategy of organization and grouping modules specialized for a defined task. Salient elements are defined by generic properties. A certain amount of ambiguities and redundancies is necessary to allow the detection of multi-scale structures. This work insists on the manipulation of complexe scenes, on usual systems. It has been applied to various type of scenes, from satellite and medical imaging to indoor and outdoor scenes. The quality of results confirm the robustness of this approach in cluttered environments.
This entry was posted in 04 - Publications and tagged combinatorial optimization, Conference presentation, Dynamic Programming, feature matching, hierarchical representation, Perceptual Grouping, reconstruction, saliency network., salient curve detection, scene representation, shape recognition. Bookmark the permalink.