Kaining Ying

same question

in the ablation study i saw the gap between the result of using GT kpts and predicted kpts

No,absoutely not.if you dont have kpts,you should have a pose esitimate model at least

maybe you should reduce the batchsize or use a bigger gpu memory machine

我是新手，我也想不懂这个问题

看了MMdetection中也是将双阶段进行一次训练，好像确实可以达到一样的效果

> In the paper, the M-net should be pre-trained. But I didn't see this step in the code. so , is the M-net should be pretrained? did you solve it？

same problem :(

mAP只有2.8这不是差了，这是由一点问题

我有，留个邮箱吧。