Visual object trapping.

A new formulation and viewpoint into the problem of object localization is discussed.We provide a discussion and definition of a different class of similarity measures adapted to such problem.The use of non-normalized histograms of quantized-color is proposed.The importance and interest of the high recall regime for tracking applications is shown. We present a video analysis task closely related to visual tracking which we call visual trapping. Classical tracking constraints and formulations are relaxed, providing a different criterion for locating an object at any time within a video sequence. The base approach essentially searches for a minimal image window such that we can guarantee to some extent that the object is inside this window. We thus consider that our acceptance criterion is that the object is inside a bounding box, instead of the minimization of position error. We then define a visual trapping algorithm that combines color-based image features with a novel distance computation in the form of an hemimetric. We discuss the advantages of this framework w.r.t previous approaches and show how our method outperforms state-of-the-art trackers for high recall regimes.