Ask Your Question

Revision history [back]

Poor Cascade Training Results

I am training a cascade to detect vehicles using Haar features. My positives:negative ratio is around 19000:12000, training at 24x24 positive images for 11 stages. However the results I obtained still yielded a lot of false positives. Could this be due to the poor preparation of data sets? I already ensured that the positive images were generally properly cropped and negative images do not contain any positive images in it.

One thing I noticed is that during my training when it was at around 8 stages and above, the false alarm rate constantly alternatve from low 0.9 to high 0.9 and will not go any lower than that, only to have the stage complete at 100 steps with a 0.9 false alarm rate. Does anyone have any idea how to fix this?