On the Use of Anchoring for Training Vision Models