weixin_39822993
weixin_39822993
2021-01-03 07:26

the IoU-Loss, Iou-Aware-Loss and final confidence in PPYOLO?

from the paper, we know that IoU-Loss and Iou-Aware-Loss (adopt BCE loss) are both used as additional loss to orig box-regression (adopt L1 loss) to improve the locate accuracy and IoU of predicted box. but in my train process, I can rarely see the decrease of IoU-Loss or Iou-Aware-Loss (as follow pic shows).

loss

so: 1. how can I reduce the IoU-Loss and Iou-Aware-Loss in PPYOLO?I've tried increasing the loss_weight of IouLoss and IouAwareLoss in ppyolo.yml and re-cluster the anchor box but didn't get promotion obviously. 2. what's the role of config item "loss_weight"? 3. while the total loss is getting decrease and AP increasing, but the final detection confidence shows a big decrease, almost around 0.5 (as follow pic shows), why? 20201104 15000 pdparams(AP50=35 8) infer result

该提问来源于开源项目:PaddlePaddle/PaddleDetection

  • 点赞
  • 回答
  • 收藏
  • 复制链接分享

9条回答