We propose and evaluate a new approach for identification of persons, based on harvesting of interest point descriptors in video sequences. By accumulating interest points on several sufficiently time-spaced images during person silhouette or face tracking within each camera, the collected interest points capture appearance variability. Our method can in particular be applied to global person re-identification in a network of cameras. We present a first experimental evaluation conducted on a publicly available set of videos in a commercial mall, with very promising inter-camera pedestrian reidentification performances (a precision of 82% for a recall of 78%). Our matching method is very fast: ~ 1/8s for re-identification of one target person among 10 previously seen persons, and a logarithmic dependence with the number of stored person models, making re-identification among hundreds of persons computationally feasible in less than ~ 1/5s second. Finally, we also present a first feasibility test for on-the-fly face recognition, with encouraging results.