Excellent breakthroughs have been achieved in various high-level computer vision tasks such as image recognition, object detection, and segmentation at the expense of growing computational costs. Recent efforts have been made to employ KD to solve this problem. Unfortunately, most previous KD methods in computer vision are designed for image classification and usually lead to trivial improvements in more challenging tasks. In this section, we delve into the design of KD for complex high-level vision by analyzing the key properties of these tasks, including detection and segmentation on images, point clouds, and multi-view images.