Task-balanced distillation for object detection

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