needs real life images, not synthetic ones.
(you can “augment” 5000 images to 15000, but not 3 to 3000 in a meaningful way)
btw, you can also try to train a HogDetector (similar performance, similar problem scope). sample here which is also easier to configure (all you need is a folder with cropped positives and another with negs)
ps: again, what about the reflection ? it looks harmful