Diagram of our co-training procedure. For the m-th image pair, an infrared/visible pair of regions of interest (ROIs) is extracted for each detection p. A score of thermal objectness is computed on the infrared ROI. If it is bigger than a threshold, the pair is injected in the training datasets as two new positive samples. If it is lower than a threshold, the pair is rejected. The new training datasets are filtered to eliminate eventual mislabeled samples. Finally, the detectors are retrained with the new data. It can be repeated n times for iterative improvements.