To design an algorithm to select robust local image features, we start by looking for what types of regions do not contribute much to visual matching and registration. First, we can filter out regions with dense textures, for example, regions with tree leaves, grasses, ripples on the water, etc. Second,many of the textureless regions are also featureless,
thus they are not good regions for extracting local features that characterize images. Examples include the sky, large patches of textureless regions on rocks or walls, etc. Third, regions which do not fall into the above two types are usually more useful for image matching and registration. Good examples are regions with man-made structures in a typical city environment. These regions usually have more consistent local gradient directions. Just think about architectural structures.