Week Eight
Over the weekend, I stayed in Princeton for the most part, and I worked. While running some tests, I found a small (but very annoying) bug which essentially meant that my performance calculations were incorrect. I fixed this error and ran the relevant tests again. I also ran tests on the second image database I had found, and as expected, the performance was much worse than with the first one. Surprisingly, WEKA didn't perform better than my classifier. I also tried implementing the deicision tree that WEKA gave me in my code, and that of course worked as well as WEKA itself, i.e.--reasonably well, but not too well.
      On Monday, I tested my classifier on webcam images I took of Maria, Rebecca, and myself, and the results were not good, but not too discouraging. I discovered that the classifier itself worked pretty well, as long as the face was actually detected. This is where Prof. Kremer's code comes in handy. I had emailed him, asking about conversion from JPEG to PNM files, and the answer I got was very helpful. I finally managed to run the code with photographs in which Visage didn't detect any faces, and on Tuesday I started looking at how I can use his code for my face detection purposes. I also had a conversation with Prof. Adam Finkelstein about grad school in graphics / vision / AI.
      I studied Prof. Kremer's code and found some numbers to output which can be useful, such as the coordinates of the box in which a face is found, as well as what I think is the nose tip. Now the problem is, 1) how do I translate his C code in Java? OR 2) how do I get the Java Native Interface running? While thinking and working on these problems, I also sent an email to all EE grads, offering them cookies in exchange for taking pictures of them. Only two showed up so far, but I hope more will later. I also downloaded Picasa to see if pictures whose contrast is modified by Picasa would allow better face detection with my current code, and this turned out to be a correct prediction. While running new tests, I read about how Java can work with MySQL commands, since we'll need some sort of database later for people's accounts for the vending machine.
      On Thursday, I kept working on learning more about JNI, and I also tried translating C into Java with some C2J program I downloaded, which only kind of worked since the C code is for Linux and I'm trying to run it on Windows since the C2J is installed on my machine, which runs Windows... On Friday, I changed the way in which scores are recorded, allowing the program to compute TP and FP scores (true and false positives) and record all these results in a file. Then I experimented with increasing and decreasing certain thresholds, in ways that correspond to a vending-machine-type and AMBER-alert-type applications of the Face Recognizer.
      Happy Independence Day to all! I was in NYC on Tuesday evening and most of Wednesday, but didn't get too see any fireworks... There were some cool ones in Princeton on Monday.

Adriana in DMP-Land :: Journal