849 Image Haarcascade

I added another 100 images from dental websites; great smiles. I also followed the exact workflow pattern described in my last blog. Just adding 100 images I got substantially better results; e.g., the lady in a yellow bikini is now recognized as smiling.

This time I got more smile hits then in my last blog. Here's the haartraining command I used:

haartraining.exe -data cascades -vec vector/facevector.vec -bg bg.txt -npos 849 -nneg 1805 -nstages 10 -mem 8192 -maxfalsealarm 0.45 -mode ALL -w 25 -h 25 -nonsym

The haartraining app seemed caught in a loop when I used the same maxfalsealarm value as last time; 0.34. 0.45 seemed to give decent results, but the session did end with the maximum false alarm rate hit, so maybe 0.44 (or lower) might have been better. I did get all 10 stages completed, so 0.45 is probably close to an optimal value.

I got my images from a google search. just like in the last blog. To see how that's done, see the following website:

https://32hertz.blogspot.com/2015/03/download-all-images-from-google-search.html

...and like before I downloaded the images with a simple Perl script (the only type I write) getFiles.pl

I made no change to facial.py as it works perfectly at extracting faces from images, and coverting them to grayacale.

These five utilities also remain unchanged:

facialSingles.py renameFiles.pl listFiles.pl renameDisparateFiles.pl convert2Grayscale.py

I hinted in my last blog that I was going to use opencv_haartraining in the near future, but I stuck with the utilities I found at:

https://www.cs.auckland.ac.nz/~m.rezaei/Tutorials/Creating_a_Cascade_of_Haar-Like_Classifiers_Step_by_Step.pdf

I actually created this new haarcascade xml file about one week ago (as of this writing), but I've been very busy lately, and just didn't have time to create a new blog. You might have already noted that this blog is a bit of a rush job, but like I said, I've been really busy. ...but I didn't want to create a new haarcascade file until I blogged about this one first. Now I feel free to proceed with 949 Image Haarcascade.