search

xingyizhou / CenterNet

Object detection, 3D detection, and pose estimation using center point detection:
MIT License
7.25k stars 1.92k forks source link

train pose estimation from strach #589

Open Dantju opened 4 years ago

Dantju commented 4 years ago

@xingyizhou ,have u train PE from strach but not finetune on OD? I did this ,and get a better result，i want to know,why u train PE finetuned on OD?

Dantju commented 4 years ago

train pose estimation from scratch, sorry get a spelling error

xingyizhou commented 4 years ago

Thanks for raising this up. I have tried and also observed better results. The new results will be updated in the next revision.

Dantju commented 4 years ago

@xingyizhou I observed a better result on PE ,but PD get a lit worse,do u have some idear to improve PD?

Dantju commented 4 years ago

@xingyizhou I find use multi_pose to train PD by setting PE's loss weight as 0，the result is worse than using ctdet to train PD ,do u have any idea?

sisrfeng commented 3 years ago

Hi, I train dla_34 and resdcn_18 on UA-DETRAC using ImageNet pretrained backbone but not loading coco-pretrained model. The AP50 is always 0 for 50+ epoch. Do you know how to solve it? Thx!

@xingyizhou I find use multi_pose to train PD by setting PE's loss weight as 0，the result is worse than using ctdet to train PD ,do u have any idea?