A Method for Visual Model Learning During Tracking

In this paper, we propose a new method for visual model learning. The algorithm learns an object representation by one-shot and adaptively extends a set of saliency filters. The filter coefficients are extracted from the environment by different views. In addition, the algorithm fuses already learned visual filters and derives new visual classifiers in order to gain generalized object concepts. We evaluate our method on tracked sequences that are resulted from a processing with a visual bottom-up attention model.