search

duzhixing / FRS

Distilling Object Detectors with Feature Richness
43 stars 6 forks source link

Experiments on detectors without FPN #4

Open ChuQiaosong opened 2 years ago

ChuQiaosong commented 2 years ago

Hi @duzhixing Thanks for your great work! I have 2 questions about FRS. Firstly the paper says FRS is more suitable for detectors with FPN moudle. I wonder if FRS can also improve perfomance on detectors without FPN. Have you done some relevant experiments? And secondly what about the performance on smaller backbones like VGG16 with FRS? By the way I am doing some experiments on Faster-RCNN with VGG16 and without FPN moudle. I used the cls_score of RPN head as the mask and performed distillation on feature map extracted by VGG16. But it seems that the performance is poor.

duzhixing commented 2 years ago

You can check this URL, https://openreview.net/forum?id=_bOfK2k_7R. . YOLOF is the latest detector without FPN module, so it is different in construction from the detector in our paper. Therefore YOLOF is suitable to evaluate the generalization of our methods. The following table shows the experiment results on YOLOF on MS COCO2017. YOLOF with ResNet50 achieves 39.3% in mAP, meanwhile surpasses the baseline by 1.8% mAP.