Descriptor Based Classification of Shapes in Terms of Style and Function

It seems that great shifts in how human beings use technology often create a push for changes to the way we divide work between human beings and technology. Chemical film has all but disappeared and almost everybody takes digital photos which they proceed to put online for easy sharing with friends and family. Together with a number of other trends which have contributed to the vast amount of online and locally stored digital photos, this has made automatic recognition of people in images an important research topic - in spite of the fact that recognition is one of the tasks generally left to human beings, since we excel at recognition. We believe that recognition of the style of a 3D object is something that is also likely to be increasingly useful in the foreseeable future. Optical scanning methodologies make the generation of 3D content more feasible than previously, and it is easy to envision digital artists wanting to compile content for a 3D scene or composite object being in need of a method for searching for an object not just of a specific function but also a specific style. The scope broadens further if we look beyond man made objects. It seems clear that, say, the various limbs of a specific human being have some commonality that separate them from those of another person. Thus, one could argue that an individual represents a style. Style in the context of biological variation is something that we explore in the work presented here. Specifically, we investigate whether we can define a style class for the teeth of a person. Unfortunately, style is subtle and we do not hope to be able to automatically extract a description of style from 3D objects. Furthermore, we avoid using explicit ways of describing style. Recognizing the style of an object based on some textual or otherwise encoded information might be a feasible approach in some cases such as, for instance, recognizing to which order a given classical greek column belongs. But, relying on explicit information about a given style would require us either to solve the above problem of automatically extracting style information from shapes or to rely on human beings to encode style - a task that we believe would be both tedious and difficult. Instead, we rely on examples in the work presented here. This requires that we have example (training) objects for each style. It also requires that we have an orthogonal class of functions, since, as we discuss below, the function of the object (what it is) clearly also has a profound impact on shape. Thus, our work can be summed up as example based classification of digital 3D shapes in both style and function categories.