Efficient Sample Mining for Object Detection

Object detectors based on the sliding window technique are usually trained in two successive steps: first, an initial classifier is trained on a population of positive samples (i.e. images of the object to detect) and negative samples randomly extracted from scenes which do not contain the object to detect. Then, the scenes are scanned with that initial classifier to enrich the initial set with negative samples incorrectly classified as positive. This bootstrapping process provides the learning algorithm with "hard" samples, which help to improve the decision boundary. Little work has been done on how to efficiently enrich the training set. While the standard bootstrapping approach densely visits the scenes, we propose to evaluate which regions of scenes can be discarded without any further computation to concentrate the search on promising areas. We apply our method to two standard object detection settings, pedestrian and face detection, and show that it provides a multi-fold speed up.