Evaluating Descriptors performances for object tracking on natural video data

In this paper, a new framework is presented for the quantitative evaluation of the performance of appearance models composed of an object descriptor and a similarity measure in the context of object tracking. The evaluation is based on natural videos, and takes advantage of existing ground-truths from object tracking benchmarks. The proposed metrics evaluate the ability of an appearance model to discriminate an object from the clutter. This allows comparing models which may use diverse kinds of descriptors or similarity measures in a principled manner. The performances measures can be global, but time-oriented performance evaluation is also presented. The insights that the proposed framework can bring on appearance models properties with respect to tracking are illustrated on natural video data.