Cool world: domain adaptation of virtual and real worlds for human detection using active learning

Image based human detection is of great interest due to its potential applications. However, even detecting non-occluded standing humans remains challenging [4]. This is not surprising due to the large variety of backgrounds (scenarios, illumination) in which humans are present, as well as their intra-class variability (pose, clothe, occlusion). Nowadays, the most relevant baseline human detector relies on a holistic human classifier that uses the so-called histograms of oriented gradients (HOG) as features, and the linear support vector machines (Lin-SVM) as learning method [2]. New methods have been developed on top of this baseline, most of them following a discriminative learning paradigm. This means that human classifiers are trained from labelled samples. Labelling is performed by a human oracle. On the one hand, the oracle must select human-free images from which negative samples can be taken, i.e., background windows. On the other hand, the oracle must draw a bounding box (BB) per each human sample of interest within non-human-free images, i.e., positive samples are image windows framing humans. In practice, this implies that a core issue as having good samples to train, relies on a subjective and tiresome manual task.