I’ve already put the subject of single object classification to bed, however, it dawned on me that there’s an entirely different approach to object classification based upon the properties of the object. We can already say what an object is, that’s not interesting anymore. Instead, what I’m proposing is to describe the object based upon its properties, in particular, texture, which is discernible through image information. Specifically, you have a dictionary of objects, and each object in the dictionary has columns of descriptors such as metallic, matted, shiny, ceramic, and you can of course include color information as well. Then, once an observation has been made, you restrict the cluster returned to those objects that satisfy the observed properties. For example, if the object observed is shiny and ceramic, then less than all objects in the dictionary will be returned. The point is not to say necessarily what the object is, though that would be interesting as well, and instead, to build what is in essence a conceptual model of the object within the machine, beyond a single label.
The applications here are numerous, from unsupervised association between objects based upon visceral information, to potentially new ways to identify objects in a scene based upon small portions of the objects. This latter approach would probably allow you to quickly identify things like leaves, birds, bark, small stones, all of which have unusual textures. Though it would also allow you to make abstract associations between scenes based upon the textures found within them, which could be the seed of real creativity for machines, since it is by definition an abstract association, that is not necessarily dependent upon subject matter or any individual object.
Information Overload 