Data storage for efficient knowledge distillation in object detectors

While the knowledge storage distillation method has proven effective in image classification, object detection is a considerably more complex task with numerous candidate bounding boxes. In this paper, we show that only a select few of these numerous candidate bounding boxes are pivotal for effective knowledge distillation. Consequently, we introduce a novel method that stores the features of important bounding boxes, selected based on their quality scores, and employs these features for learning using quality score-based masks. Through experiments, we demonstrate that our method significantly reduces computational overhead while preserving performance.