The problem of dense matching in wide baseline stereo is considered. It is assumed that the scene is formed by piecewise-smooth, Lambertian, textured surfaces and that rectified images of the scene are available. We present a novel dense matching algorithm that accounts for local affine distortion and propagates the best matching affine parameters on each surface until a surface discontinuity is reached. Disparity maps and projective reconstructions for real-world wide baseline stereo images are shown.