A problem with these window-scanning approaches is that the idea of training a separate classifier for each object in the world (and even different classifiers for different poses) just doesn't scale with the large number of objects present in the world.


Surely one does not want 1 million classifiers being independently applied to each window in an image.