Detection of close objects using CascadeObjectDetector

I am trying to train a cascade object detector to work on non-human faces, but it doesn't detect them, especially when they are large (as in they take up a large part of the image). In my training set, I have around 50-something positive images per angle of face (front, 45 degrees, and profile. I have around 40 negative images, which are all backgrounds that do not contain positive instances. I've been running

trainCascadeObjectDetector('Front.xml', positiveimagestuctarray, negativeimagedirectory, 'FalseAlarmRate', 0.2, 'FeatureType', 'Haar', 'ObjectTrainingSize', [150,150]

to train the detector, but it fails after around 4 stages. When I try running it, it only detects small parts of the face, such as the eyes or random parts of the skin. Am I wrong to think that the object training size should be larger? When I try training at larger training sizes, it crashes with an out of memory error.

Answer 1

[Dima Lisin]

The main issue here is that you need more training data. 50 positive samples and 40 negative images is very little, which is why training quits after 4 stages. It is simply running out of data. For a good detector you need hundreds, or better yet, thousands of positive samples, and a similar number of negative images.

Second issue: having a mix of front and profile faces in your training set is a bad idea, because they look very different. Typically you have to train separate detectors for front and profile.