Distilling object detectors with efficient logit mimicking and mask-guided feature imitation

Xin Lu, Yichao Cao, Shikun Chen, Weixuan Li, Xin Zhou, Xiaobo Lu

Abstract: Highlights•Distillation via classification logits can benefit one-class object detection.•Logit mimicking for localization mitigates missed and false detections.•Teacher sample selection enhances knowledge transmission efficiency.•A soft metric and a loss-guided mask are more suitable for feature imitation.