Meeting Date: 16.08.1999
What I did last week: Shaun helped me port the SKETCHUP routines to unix, which basically involved cutting out all the DOS-specifics. He got it to the point where we could recognize Gray-map pgm-files. I then continued on my own removing all the edge-detection so that I could produce the edge information myself. I then spent the week-end figuring out, how the recognition actually works (i.e. why the training-files had exactly 66 discontinuities), which also involved getting more aquainted with size-functions. On Monday I sat down with Shaun again and we discussed the data-files. He had a great idea about the format of these files, which led me to writing my own application using the SKETCHUP recognition system and I even compiled my own set of training data. The recognition actually works, but is far from satisfactory. The gestures I used were the standard ones, which proove to be too simple for this recognizer, which works on graphs containing shape-informations. The existing gestures are not very rich in topological information and thus appear very similar to the recognizer.
What I'll do this week: I want to compile another training set with more complex, yet easy to draw gestures and see if the recognition rate improves. I also want to do some testing to correlate number of vertices to recognition rate and similar things (i.e. do some real research). I will decide about whether to try to improve their algorithms or not. So far they are not using their connectivity information to the full extent, because they assume a simple closed shape without intersections as their edge-input. My program can supply more topological information that that. It would be interesting to see, how this influences the recognition quality.