A cluster-based strategy for active learning of RGB-D object detectors

We present a method to detect human body parts in depth images that is based on an active learning strategy. Our aim is to built an accurate classifier using a reduced number of labeled samples in order to minimize the training computational cost as well as the image labeling cost. The active learning strategy is based on exploiting the training data distribution by sampling from a cluster-based representation of the dataset. We show that this strategy allows a significant reduction of the number of samples required to train a high performance classifier. We validate our approach on two different scenarios: the detection of human heads of people lying in a bed and the detection of human heads from a ceiling camera.