We have a project that will identify a logo from an image. We initially used a Haar classifier, but training a Haar classifier take lots of time (4 days per logo on our Core i5 machine). To train it for 300+ logos will take a lot of time (we do not have any high performance computers). So, we have decided to move to a HOG based object detector, hoping that its training will take significantly less time.

Does anyone have an idea how much time HOG descriptor training takes? We would be training on the about 100 positive and 100 negative 600x800 pixel images per logo (on a machine with a Core i5 processor).

Answers:

Unanswerable, depends on the number of bins and other implementation details. Probably also on the content of the images. Don't expect it to be super fast with 60k images though. If I were you I would seriously consider downscaling the images, 600x800 is much larger than what you need for recognition. 150x200 should still be recognizable, but all the computations would be 16x faster.