search

HRNet / HigherHRNet-Human-Pose-Estimation

This is an official implementation of our CVPR 2020 paper "HigherHRNet: Scale-Aware Representation Learning for Bottom-Up Human Pose Estimation" (https://arxiv.org/abs/1908.10357)
MIT License
1.35k stars 272 forks source link

Group #7

Open BobDL opened 5 years ago

BobDL commented 5 years ago

Hi, Thanks for your great job and the open source. I test the pretrained(w32-512)model on the coco set. And for most the results are good. But part of them, the keypoints are grouped in obviously wrong way. Is there any opportunity to improve the group?

Thanks image

bowenc0221 commented 5 years ago

As far as I know, it is hard to improve this kind of error simply using post-processing. Probably you can try different grouping methods other than AE.

BobDL commented 5 years ago

As far as I know, it is hard to improve this kind of error simply using post-processing. Probably you can try different grouping methods other than AE.

Thanks for your reply. Have you tried the paf grouping method?

bowenc0221 commented 5 years ago

No, we haven't. We only tried AE in our work since our focus is not on the grouping part. But using better grouping should definitely help.

hellojialee commented 4 years ago

@BobDL I have tried the PAF-like (proposed by OpenPose) grouping method and pre-trained HRNet-w48 with Focal L2 loss based on this repo Simple Pose but only get 63 AP on COCO test-dev dataset for now. Multi-scale inference is used during evaluation.

massir-rpi commented 4 years ago

Hi @BobDL I am encountering a similar issue. Did you find a good way to improve grouping?