Label smoothing is widely used in deep neural networks for multi-class\nclassification. While it enhances model generalization and reduces\noverconfidence by aiming to lower the probability for the predicted class, it\ndistorts the predicted probabilities of other classes resulting in poor\nclass-wise calibration. Another method for enhancing model generalization is\nself-distillation where the predictions of a teacher network trained with\none-hot labels are used as the target for training a student network. We take\ninspiration from both label smoothing and self-distillation and propose two\nnovel instance-based label smoothing approaches, where a teacher network\ntrained with hard one-hot labels is used to determine the amount of per class\nsmoothness applied to each instance. The assigned smoothing factor is\nnon-uniformly distributed along with the classes according to their similarity\nwith the actual class. Our methods show better generalization and calibration\nover standard label smoothing on various deep neural architectures and image\nclassification datasets.\n