search

Megvii-BaseDetection / BorderDet

BorderDet: Border Feature for Dense Object Detection(ECCV2020 Oral)
Apache License 2.0
430 stars 63 forks source link

I have the same question. #46

Closed StetchPlane closed 3 years ago

StetchPlane commented 3 years ago

https://github.com/Megvii-BaseDetection/BorderDet/issues/15#issuecomment-706919567

How can I solve this problem? I'm looking forward to your corrections.

Maycbj commented 3 years ago

I'm sorry that I reply to you so late. We have tried to use 9 anchors to train our method based on RetinaNet, which is similar to the anchor-free detectors with BorderAlign. But it is computationally intensive. So we tried different strategies to reduce the amount of calculation. 1) choose the highest score to train 2) choose the highest IoU to train 3) randomly choose an anchor to train Strategy 3) yields the best performance in our experiments, owing to we supervise them equally to overcome overfitting. Besides, I suggest that BorderDet based on Anchor-Free detector may be a better choice for you to get better performance.