Collaborative training of far infrared and visible models for human detection

This paper is about the collaborative training of a far infrared and a visible spectrum human detector; the idea is to use the strengths of one detector to fill the weaknesses of the other detector and vice versa. At first infrared and visible human detectors are pre-trained using initial training datasets. Then, the detectors are used to collect as many detections as possible. The validity of each detection is tested using a low-level criteria based on an objectness measure. New training data are generated in a coupled way based on these detections and thus reinforce both the infrared and the visible human detectors in the same time. In this paper, we showed that this semi-supervised approach can significantly improve the performance of the detectors. This approach is a good solution to generate infrared training data, this kind of data being rarely available in the community.