In my last blog I said I said my next haarcascade file would incorporate 2000 images. As usual, things got pretty busy for me after
uploading my last blog in this series; work at the Organic Forest, new projects around the house, One of mys disk drives crashed (nothing I did to save it worked,
but I wasted plenty of time trying - my last backup of that drive vanished, and it took a while to rebuild the drive). During all that time I received e-mail from
a former classmate (we took an html5 development course together at Atlantic Technical College), and he asked me how the emotions program was going. I was only up
to 1300 images when I received the e-mail, so I downloaded another 200 images (and another 1000 negative images) and created a 1500 image haarcascade file.
I created the file with the following command:
haartraining.exe -data cascades -vec vector/facevector.vec -bg bg.txt -npos 1500 -nneg 3283 -nstages 12 -mem 8192 -maxfalsealarm 0.46202 -mode ALL -w 25 -h 25 -nonsym
I expected continual progress; i.e., I expected the value of maxfalsealarm to go continually down, but the first time I ran this batch, I used maxfalsealarm value of 0.392; the file
was overtrained (no smiles were detected). In fact, I'm not getting any more smile detection hits then I did on my last run with 1000 images. The smile detection haarcascade file that
comes with OpenCV, haarcascade_smile.xml, doesn't find any more smile then my 1000 image file. The video file I keep using is a montage of smiling actresses that I downloaded from
youtube.com, and the girls are facing in all sorts of directions. The Viola Jones algorithm is built to detect frontal face images, so maybe I should quit using this video, and go back
to using my own webcam. However, I'm a diehard, so I'll throw in at least a couple of thousand more images before starting my next emotions detection haarcascade file.
I started this blog series with a song called, "No Feelings", by the Handsome Furs. I used several more songs that referred to a lack of emotions or some similar type of condition. Pictures of Matchstick Men refers to the way haarcascade files break up images into component vectors. "Pictures at an Exhibition" is about pictures, "Pictures of Lily" is about pictures. The rest of the songs have been about smiles. I used Words, by the Bee Gees, because it starts with the lines "Smile an everlasting smile. A smile can bring you near to me."; sticking with the smile theme.