Let me introduce myself first. My name is Daniel Kirsch and I am a math student from Münster, Germany. The last three days have been very interesting. I launched (the second version of) a Web application that I developed over the last few weeks and it has had way more publicity than I would have ever imagined. Yesterday when Detexify was on reddit I had trouble to keep the service even usable.
I have gotten many nice emails and many of the comments on hacker news or reddit and a lot of tweets were really flattering. It seems like I have built a really useful piece of software.
Now, the funny thing is, I really don’t know how well Detexify actually performs. I don’t use LaTeX right now (although will definitely have to in a few weeks/months). Of course I have done some tests und it doesn’t look too bad. But with all the training data you provided it should be possible to actually measure the performance.
The idea for Detexify came up during a lunch with a fellow student (Philipp Kühl). We discussed about what ideas are worth implementing and I argued that solving problems of your own often solves the same problem for people in similar situations. He gave a LaTeX course at that time and remarked that he often encountered the problem of not knowing the command for he certain mathematical symbol he had in mind. So this was not my problem at the time but being a web2.0 nerd I thought “That would make a neat web app!”. And I went to work…
Until then I had no experience with handwritten OCR or any kind of pattern recognition and the hardest part was coming to grips with the topic and learning how things could possibly work out. The actual implementation is so simple that its almost embarrassing but I will explain the inner workings of Detexify in a later post. Whoever is brave enough to find out by himself can dive straight into the code though.
For the past days I have mainly worked on performance optimizations but I also got a few requests to add missing symbols. I will wait a few more days until the traffic on Detexify stabilizes and then add new symbols in one big update. Of course they will need training but my impression is, that will take almost no time. In the last four days you have provided over 23.000 training samples. You have had a big share in making Detexify what it is.