Knowledge distillation methods provide task knowledge from complex teacher networks to compact student networks, aiming to improve the performance of the student network. Usually the knowledge is transferred through channel-wise feature mapping, from both final outputs and intermediate features of the student and teacher networks. When the number of channels in the feature maps of the student and teacher networks differs, previous methods predominantly employ feature transformation to align these channels. We propose a novel approach to discretely align intermediate feature channels by selecting discriminative features from the teacher network feature maps, rather than transforming them. Our new mapping mechanism achieves effectiveness by capturing instance-wise feature importance. Meanwhile, it is structurally simple and computationally cost-effective. Extensive experiments on classification and object detection tasks, across network architectures of similar and dissimilar design styles, clearly demonstrate the effectiveness and strong adaptability of our approach.