movenet.pytorch icon indicating copy to clipboard operation
movenet.pytorch copied to clipboard

Published 20 hours ago • verified publisher icon fire717

Sota result or just for fun？

Open 2050airobert opened this issue 3 years ago • 2 comments

2050airobert avatar Jun 11 '22 09:06 2050airobert

这些问题readme和谷歌博客其实都有答案, 就不能先看看再提问吗？

  1. 原始模型因为有大量非开源内部数据训练(Evaluations on the Active validation dataset show a significant performance boost relative to identical architectures trained using only COCO. )，在没有同样数据训练的情况下，你没办法比较复现的模型效果是否一致
  2. movenet本身追求的是轻量级，是嵌入式设备的实时性，因此精度上肯定比不上sota模型（比如bottom-up的openpose、top-bottom的hrnet），只能说在有足够、高质量的数据下还是能取得不错的精度
  3. In readme part3：Surely this is a muti-task learning. So add some loss to learn together may improve the performence. (Such as BoneLoss which I have added.)

fire717 avatar Jun 11 '22 14:06 fire717

test

2050airobert avatar Jun 13 '22 01:06 2050airobert