I think the previous failure of filter bank approach was caused by alignment of those windows. In this post I am using a fine 2 by 2 window, overlap by 7 pixels, to slide through the input image(actual window size is 16 by 16, 2 + 7*2).
It rejects those negative objects pretty well but also does not fire on the positive bees. Again, I may need a bigger sample size to get an ideal result.
UCSD CSE190a Haili Wang