Finally, I would like to end with a question that came up when we
began to examine our initial results. When we looked at the images we
generated with our filtered silhouette
maps, we noticed that the images get a cartoon-like quality to them.
For example, the teddy bear head here has been filtered like we wanted,
but the result looks more like a cartoon
than the original image did. Why is that?
A little thought shed some light on the problem. Our input image has some given frequency distribution, which I show here on the left. Because we generate the color samples for our representation by averaging down the original image, we can think of the process as acting as a low pass filter. In addition, we artificially embed the high-frequency edge information by adding the silhouette map, and thus add a sort of high-pass filter to the process.
So when we pass our input through a filter with low and high-frequencies but no mid-range, we get an output that looks like this on the right. Our vision system automatically recognizes something with low and high frequencies but no mid-range as a cartoon, because if you think about it, cartoons are typically made up of black outlines (which provide the high-frequencies) around regions of relatively constant color (the low frequencies).
This observation got us thinking about interesting potential future work...