Haar Training with object Blacked out in Negative Samples

Question

I am wondering if there is any downside to blacking out the target object when it appears in negative samples during Haar Cascades training.

Basically, If i want to detect a human nose. Can I just get lots of face images and black out the nose for the negative samples. I would then get lots of cropped pictures of the nose for positive samples.

I was wondering if there is a downside or problem with having lots of negative samples with the same non-nose feature e.g. a black rectangle. Is this going to train the cascade to expect a black rectangle in the negatives....????

Sorry if this questions seems naive but I have not done haar training before and the answer might help others who are investigating the training process...

Answer 1

I have had good success doing this with my training sets. I have modified the OpenCV_Annotation application to present a fixed rectangle that moves with the mouse, once the space bar is pressed the coordinates are written to the positives.txt file and a copy of the image is saved into a pre-defined negatives subdirectory after graying out the rectangle. I have seen no training problems or performance issues although I use other negatives as well as the blurred out images. The application uses a gray background to overwrite the identified area instead of just blacking it out. I am unsure if that would make a difference in the negative.