Abstract: In this work, we show that deep learning-based re-identification (Re-ID) models, albeit trained only with a Re-ID objective (i.e. if two samples belong to the same identity), encode ...