The alternative to window correlation is to process each 2D image independently to detect significant features. Image features should be chosen to provide a consistent representation of what is physically present in an image. The detected features in the image array should arise from physical features in the real world (shadows, surface markings, surface edges, etc). Also, if the same object is present in two different images then a representation (constructed from image features) of the object in one image must closely resemble the representation in the other. Image features are gauged on how accurately they perform this task, and trying to match image features would prove virtually impossible without consistent feature detection.

The most popular image features detected in natural (real world) images are edges ; stereo vision systems have been constructed in which either single edge pixels or extended boundary features are matched between images. Edges are detected as sharp pixel intensity changes occurring in an image due to discontinuities in viewed surfaces, shadows, surface markings and so on. Edges also satisfy the uniqueness constraint, which has become virtually a matching program standard. Matching across images can only take place between unique image features, so no one feature in an image can match more than one feature in another image.