Natural images remain poorly understood, except for some phenomenology that appears to be common to virtually all "natural images" such as the self-similar (fractal) spatial structure. However, this "universality" it is not very informative. More specific image structure depends upon the physical causes of the radiance function ("light field") and on the imaging process. The latter is well, the former hardly understood. The light field is due to both the "primary sources" as well as to multiple scattering (mostly by rough surfaces) in the scene. Either component may dominate, depending on location in the scene. The optical interactions depend mainly upon geometrical optics processes (interposition, vignetting, attitude effect, scattering by rough surfaces ...) on many different scales. Only the molecular scale is well understood (standard physics). Most substances from the natural environment are non-homogeneous, have rough boundaries, and are quite different on different scales. Descriptions must span the range from leaf, leaf cluster, foliage, tree top, forest, wooded area, ... for instance, depending upon distance and resolution. The appearances of surfaces objects depends upon the light field and the light field depends upon the scene (the objects).
During the 1980's and '90's computer vision almost singularly focussed on (chrono-)geometrical issues (the WHEN, WHERE questions). The WHAT questions have been neglected and will become an important focus in computer vision/image processing/computer graphics for the next decades. Important topics that might be addressed are estimation of the light field (in addition to/together with the geometry), segmentation of natural objects, classification of materials on the basis of (illumination dependent) texture, (estimated) BRDF and color, "color constancy", classification of the setting ("landscape", "cityscape", "office environment", ..), and so forth.