Revision history [back]

When you train a detector there is usually no problem to get a large number of negative examples. Even in case when the camera is fixed I would highly recommend to use a large number of negative examples. It is true that some part of the background will be always the same (or similar, dependent on illumination) but soms of the background will vary and you don't know how in advance. For example if you were to detect heads from top view, heads will be positive examples, but shoulders and legs will be negative examples. Detector shouldn't fire on those as well. When you have large negative set it will lead to lower false positive rate.