2000 Image Haarcascade

In my last blog I said I said my next haarcascade file would incorporate 2000 images; this is it. I have the flu right now, so I don't feel like going into a whole lot of detail about what I did when creating this file, but I do want to point something out that was getting obvious to early on when creating these xml files; a more "accurate" file isn't necessarily better at getting more smile hits. By "accurate", I mean a file that has a low value for maxfalsealarm, and presumably a high number of stages. I just discarded my latest iteration of this 2000 image file because it was getting less smile hits then the previous version. I created the file used to create the video on this web page with the following command:

haartraining.exe -data cascades -vec vector/facevector.vec -bg bg.txt -npos 2000 -nneg 4034 -nstages 10 -mem 8192 -maxfalsealarm 0.38 -mode ALL -w 25 -h 25 -nonsym

I only got 9 stages completed before the program terminated with a max error message. Naturally I tried to improve my results by going to -maxfalsealarm 0.37, which yielded the full 10 stages, but still terminated early with a max error achieved message. When I tried using the 10 stage file in my python script, it didn't get as many smile hits as the 9 stage file. If the 9 stage file was getting a bunch of bogus hits, then I'd understand why the 10 stage file was getting less hits, but both files are only getting hits on actual smiles. (It just occurred to me that if somebody was looking at this web page as their introduction to my emotions program, it wouldn't make a lot of sense. This series starts with the No Feelings web page).

Angelina Jolie is smiling in the video, and my program gets a hit. The next several actresses are smiling too, but they don't get smiling hits. The Viola Jones algorithm was created to just detect faces (and anything else, but that was the original intent of the algorithm) only from a full frontal view. In fact, the algorithm picks up smiles from heads that are slightly rotated, and some other positions, but not by much. That explains some of the missing hits in this video, but some of the missing hits seem full frontal enough to be hits; something I need to look into. I'm getting hits on a lot of smiles that I was missing with earlier versions of my smiles haarcascade file, so overall I've made progress.

In my last blog I said that I was sticking with the smile theme for the background music for the web pages of this series. Jerry Lewis has died since the last blog, so in memory of the comedical actor, I'm using his theme song, Smile.