Finding an OSPA based object detector by aweakly supervised technique

The design of multitarget tracking procedures includes, as the most time consuming steps, the definition of the objective class and the formulation of the detection criteria. In this paper we investigate a solution toward an intuitive way for implementing a detector for any ad-hoc application. We capitalize on the OSPA metric to discriminate between the semantic object class of interest and other look-alike classes starting from a short number of unlabeled markers. We propose an illustrative algorithm with a toy example, then we apply it to two real images, the first acquired by SEVIRI, the second by MERIS. In the first case we discriminate between lakes, sea and look-alike clouds, in the other between ground and sea ice. We show how semantic classes with very similar spectral properties can be separated even in the presence of uncertainties or errors in the ground truth.