| A wide range of problems in computer vision is addressed by analyzing local evidence and then integrating the evidence over large regions. The first problem addressed is that of interpreting oriented patterns. At a local level, these patterns can be described by specifying whether or not there is a dominant orientation and if so, what the dominant orientation is. On a larger scale, the orientations can be combined into a coordinate system which greatly simplifies further analysis of the pattern. The second problem is finding salient contours such as edges, lines or subjective contours in images. An optimization technique is developed which allows local evidence for a contour to be integrated over large distances by elastic contour models called snakes. A user can guide a snake toward a particular feature by interactively exerting forces that push it towards the desired local minimum. A continuation method based on a spatial scale parameter allows the snake to find good local minima from far away. The result is a power-assist in image interpretation. The user places the snake near a salient contour and the minimization pulls it in the rest of the way. Once a snake has come to equilibrium on a contour, it can track it over a wide range of motion. The snake paradigm is also extended to higher-dimensional problems such as image-matching and symmetric object reconstruction. For image matching, an elastic sheet is subjected to a force derived from image cross-correlations. For symmetric object reconstruction, an elastic tube is given a preference for axial symmetry. The resulting symmetry-seeking model is subjected to forces that pull its silhouette towards observed image edges. The model is able to reconstruct the shape and motion of a bending finger from a motion sequence of stereo images. |
