Haar Classifier Training Program

If you haven't used Processing to power through some small application that you need, I highly recommend it. (I just now remembered I still haven't posted about my poster in the conference, and given that it's officially published, I can officially talk about at least the extent to what we've done so far. Next post will be on that.) In any case, as part of the research, we're now switching to an OpenCV implementation, which means training a Haar Classifier. Now this is a task that is...a pain. It involves ideally thousands of images, and if you want the best accuracy, you should hand-annotate where your object is in the images.

To avoid going crazy and needing to destroy everything in the lab, I wrote a short app in processing to create the Haar file. I'll probably post a link to the code after I've completed the training on our images and verified that it doesn't start eating memory like no-one's business for large samples, since I had tried so very hard to avoid writing this app myself. Still, using Processing, it took all of 30 minutes to produce something that will be useful. I could've spent more time with the interface, and if I decide to release it, I very well may modify the code to have a nicer I/F and support multiple occurrences in the same image, but for now, this little app is sufficient for my purposes.

For the record, there's something utterly satisfying about completing a program in under an hour. :)

UbiComp 2010: Paper of MY Conference

So while this paper was not nominated for best paper, and should not have been, it did quite a bit for me personally in terms of considering structure for my own future thesis.

Andrea Grimes, Vasudhara Kantroo and Rebecca E. Grinter (2010). Let's Play! Mobile Health Games for Adults. Proc. of UBICOMP 2010. (Link is to Ms.Grimes website, and you can find the full paper on the publications page.)

Let me sum up the gist of the idea, rather than going into the details of the paper, as there are some fundamental problems with the study. The gist is that a casual game might be better suited for games that aim to improve behavior. The behavior here was based on healthy eating, and the research targeted African American individuals because of the high incidence of diabetes among that population. The game was designed with food items that such folks would find in their neighborhoods, including soul food type items, sandwiches, etc.

The critical part for me was the introduction of the casual game premise. I had been structuring my thesis idea around epistemic games, thinking only in terms of long story arc type things, and listening to this proposal has given me an alternate idea. I'm definitely going to need to start with an exploratory study to examine which of those two paths is going to be the better choice for me, I think. Additionally, discussing the shape of the study and what might have been a better question for the study format has helped me think about how to structure my own studies to avoid those types of critiques.

I'm likely going to blog about a few more of the papers that were interesting or that I discussed in a critical fashion with Leilah as well.