Mapping COCO keypoints to Human36 format

[aitikgupta]

I understand that Stacked Hourglass predictions (MPII format) are permuted to Human36 format. To infer, as mentioned here in issue #15, it is possible to pre-process annotations in-the-wild to Human36 ground-truth 2D format.

I was wondering if the same permuting work has been done for COCO keypoints format? Any pointers are appreciated. Thanks!

[garyzhao]

Hi @aitikgupta ,

Thanks for your interest in our work!

One possible solution for mapping COCO to H36M format can be found here: https://github.com/JimmySuen/integral-human-pose/blob/master/pytorch_projects/common_pytorch/dataset/hm36.py#L73

Or I think you can reduce all the numbers of key points to 13 and then mapping them.

Best, Long

[HDYYZDN]

@garyzhao ，您好， 请您推荐一个复现Stacked Hourglass Networks比较好的代码 ，非常期待您的回复

[garyzhao]

@garyzhao ，您好， 请您推荐一个复现Stacked Hourglass Networks比较好的代码 ，非常期待您的回复

Hi @HDYYZDN ,

You can check this one: https://github.com/bearpaw/pytorch-pose

Best, Long