Task-balanced distillation for object detection. (May 2023)

• Record Type: Journal Article
• Authors: Tang, Ruining; Liu, Zhenyu; Li, Yangguang; Song, Yiguo; Liu, Hui; Wang, Qide; Shao, Jing; Duan, Guifang; Tan, Jianrong
• Is Part Of: Pattern recognition. Volume 137(2023)
• Publication Date: 2023-05
• Subjects: Object detection -- Knowledge distillation -- Computer vision; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2023.109320
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A novel HD component is proposed to assist the generation of harmonious predictions for the student. A novel TFD module is devised to mimic the intermediate features by balancing the contributions of two sub-tasks. Abundant experiments are conducted. The proposed method is easily plugged in and achieves SOTA performance. Abstract: Mainstream object detectors are commonly constituted of two sub-tasks, including classification and regression tasks, implemented by two parallel heads. This classic design paradigm inevitably leads to inconsistent spatial distributions between classification score and localization quality (IOU). Therefore, this paper alleviates this misalignment in the view of knowledge distillation. First, we observe that the massive teacher achieves a higher proportion of harmonious predictions than the lightweight student. Based on this intriguing observation, a novel Harmony Score (HS) is devised to estimate the alignment of classification and regression qualities. HS models the relationship between two sub-tasks and is seen as prior knowledge to promote harmonious predictions for the student. Second, this spatial misalignment will result in inharmonious region selection when distilling features. To alleviate this problem, a novel Task-decoupled Feature Distillation (TFD) is proposed by flexibly balancing the contributions of classification and regression tasks. Eventually, HD and TFD constitute the proposed method, named Task-Balanced Distillation (TBD). Extensive experiments demonstrate the considerable potential and generalization of the proposed method. Notably, when equipped with TBD, the performances of RetinaNet-R18/RetinaNet-R50/Faster-RCNN-R18 can be boosted from 33.2/37.4/34.5 to 37.3/41.2/37.7, outperforming the recent KD-based methods like FRS, FGD, and MGD.