Abstract

Texture density depends on both the orientation and the depth of the textured plane in view. Previous approaches stated that because of this fact (i.e., density depends on both scaling and foreshortening), density cannot be used to recover surface orientation under perspective projection. In this paper we prove that these two effects (scaling and foreshortening) can be separated and so texture density can be used to uniquely recover surface orientation. We present algorithms that are based on strong (texels) and weak (edges) segmentation. Experimental results on natural images, based on the Gibsonian uniform density assumption, are very good; these images include grass fields, gravel paths, brick walls, aerial photographs of towns or parking lots, ocean waves, man-made artificial texture (cloth, carpet, etc.). Our algorithms first preprocess the image to find texels (and if this is not possible, to find edges) and then using the assumption that the texture-elements are uniformly distributed on the world plane (Gibson), they recover its orientation. An extension of our theory to curved surfaces is also discussed.

© 1985 Optical Society of America