We solve the stereo correspondence problem by using extended edge contours as a source of primitives for structural stereopsis, as opposed to traditional point-based algorithms. Since an extended image feature conveys more information than a single point, its spatial and photometric behavior can be exploited to advantage. There are also fewer features to match, yielding a smaller combinatorial problem. The structural approach permits greater use of spatial relational constraints, eliminating the course-to-fine tracking of point-based algorithms. Solving the correspondence problem at this level requires only an approximate (probabilistic) characterization ofthe image-to-image structural distortion and does not require detailed knowledge of the epipolar geometry. This approach offers potential in automatically rectifying stereo images for which the camera geometry is known only vaguely a priori, the situation one encounters in aerial photography. Similarly, these techniques may be used to infer transformations between images of the same scene recorded by different sensors, at different times, but which are not precisely colocated in space, such as often occurs in aerial photogrammetry and remote sensing, as well as in mobile robotic navigation. We present experimental results in matching and calibration on real images using Laplacian-of-Gaussian (LoG) contour fragments as primitives in structural stereopsis.