Result is bad!

[herochen7372]

It is effective on the photos you provided, but the pictures I collected myself, even if the movements are similar, the effect is very poor. What is the reason, please.

[mks0601]

Hi, this is expected as image appearances of interhand dataset is far from those of in-the-wild images. Please check my recent work for this
https://github.com/facebookresearch/InterWild

[Javacr]

Hi, this is expected as image appearances of interhand dataset is far from those of in-the-wild images. Please check my recent work for this https://github.com/facebookresearch/InterWild

If I'm not mistaken, the reason is that @herochen7372 don not edit bbox, so the model cannot find joint in cropped image.

[herochen7372]

Hi, this is expected as image appearances of interhand dataset is far from those of in-the-wild images. Please check my recent work for this https://github.com/facebookresearch/InterWild

If I'm not mistaken, the reason is that @herochen7372 don not edit bbox, so the model cannot find joint in cropped image.

Thanks for the reply, you're right.

[mks0601]

You can use InterWild (https://github.com/facebookresearch/InterWild), which has its own hand detector, so you don't need to set bbox in advance.