paper https://openaccess.thecvf.com/content/WACV2021/papers/Chawla_Data-Free_Knowledge_Distillation_for_Object_Detection_WACV_2021_paper.pdf

code https://github.com/NVlabs/DIODE

Summary

• DIODE: pre-trained detection model로 부터 model inversion을 이용한 synthesize images 의 프레임워크
• a data-free mimic learning approach to distill the knowledge from a teacher to a student on the synthesized images for object detection
• Pre-trained: Yolo-V3, Student: new Yolo-V3
• Synthesized image들을 이용한 detection이 same domain proxy dataset 과 비슷한 성능까지 가능

• 즉, generated datasets을 이용한 knowledge distillation

Experiments

• DIODE는 Yolo-V3 이외의 다른 detector에서도 잘 적용됨

Fig 2 vs Fig 5(on SSD 300)