1000 Image Haarcascade

In my last blog I said I said my next haarcascade file would incorporate 949 image, but in my intro monologue I stated that was changeing to 1000 images. This blog is about the 1000 image file. I created the file with the following command:

haartraining.exe -data cascades -vec vector/facevector.vec -bg bg.txt -npos 1019 -nneg 2105 -nstages 10 -mem 8192 -maxfalsealarm 0.392 -mode ALL -w 25 -h 25 -nonsym

You can see I downloaded more smile images then I expected, and created a 1019 image file. The training session ended with a message saying I had met the maximum error rate, so my file might not contain all of the images I downloaded. Maybe that's why the lady in a yellow bikini is no longer getting a smiling hit. However, some video frames are now getting smile hits, that weren't getting hits before. So as of this writing I'm running another training session with a maxfalsealarm rate of 0.391, and maybe that will give me a few more smile hits (still using the same video of smiling lady celebrities).

Haarcascade Smiles File

Fast forward, and the haarcascade file with a maxfalsealarm value of 0.391 is finished; still ended with max errors achieved. The lady in a yellow bikini still isn't showing up as a smile, so as of this writing I'm trying out 0.3905 (I tried 0.39 and it got caught in a loop). I mentioned in an earlier blog that the haarTrining.exe only provides 6 digits of accuracy for maxfalsealarm (using gentle adaboost), but that means I might need to use .390001 (if that still terminates early, and I know .39 is too low for 1000 images and 10 stages, then I'd have to increase the number of stages - that risks overtraining, and not getting any smile hits at all). Here's the video from using a haarTraining command of:

haartraining.exe -data cascades -vec vector/facevector.vec -bg bg.txt -npos 1019 -nneg 2105 -nstages 10 -mem 8192 -maxfalsealarm 0.391 -mode ALL -w 25 -h 25 -nonsym

I'll show the video created with the maxfalsealarm = 0.3905 in my next blog. I'm also going to add images from the video used for this page (and the last several blogs) and modify facial.py; the program that creates images of just the head from my downloaded images (and converts them to gray scale). The program video_streamer14.py (the smile detection program) looks for head positions that aren't quite frontal, head on images. I still have facial.py looking just at frontal, head on images; I'm going to modify that to include slightly off kilter heads. However. my next blog won't come out untill I have accumulated 2000 smiling head images (as promised in my homepage intro, made right after the last blog).