This work argues that repeatable regions are not necessarily discriminative and can therefore lead to select suboptimal keypoints, and proposes to jointly learn keypoint detection and description together with a predictor of the local descriptor discriminativeness.