Learning Person Detectors from Multiple Cameras

Recently, object detection by combining information from multiple cameras has become popular. We adopt these ideas for on-line learning of object detectors from multiple cameras. Assuming that each camera corresponds to a classifier, we extend the original co-training approach to visual learning by exploring geometrical information extracted from the visual input. In this way, new, very valuable training samples can be generated, that would not be obtained otherwise, allowing us to train increasingly better classifiers without hand labeling. The approach, hence not limited to this application, is demonstrated on a person detection task for two different challenging scenarios. In addition to a performance and stability analysis we show that the proposed method outperforms existing approaches in terms of accuracy and recall; even using far less labeled data and much simpler descriptors.