Geometric Representation and Grouping of Image Curves

摘要: Perceptual organisation, while acknowledged to be important computer vision, is often based on heuristics rather than geometrical methods. This thesis highlights the role which geometry plays in perceptual organisation including geometric representation of image curves and grouping geometrically-related curves. First, via B-splines considered. Current methods are deficient because they do not fully take into account main issues spline fitting include selecting number distribution control points, parameterisation sampling data points. By addressing these with a combination active contours, minimum description length principle structural-mechanics-inspired point insertion strategy, superior results obtained. Having provided for curves, related by affine symmetry transformations investigated. Means quantifying local contour points uncertainty analysis differential properties provided, together spatial symmetries method recovering global configuration. It also shown how correspondence hypotheses between can localised using curve neighbourhood around correspondences, quantitative measures both localisability accuracy correspondences proposed. Salient accurate highly localisable thus used as ‘seeds’ remainder Finally, integrated representations grouped contours developed exploited tracking different views. Different restrictions imposed deformation geometrically-coupled conditions, symmetric, planar epipolar constraints. that robustness noise improved when deformations subjected tighter The algorithms were successfully tested real images sequences contain background clutter occlusion.