vchoutas / expose

ExPose - EXpressive POse and Shape rEgression
Other
606 stars 99 forks source link

RCNN or OPENPOSE? #23

Open lll-gen opened 3 years ago

lll-gen commented 3 years ago

Hi,it's a great job. May I ask when you train your model,you use openpose or rcnn to crop Hand and Head?and which you think is the best way?

ortegatron commented 3 years ago

Hi, keypointrcnn_resnet50_fpn is used for body detection. I don't know why author decided to use this, but probably is a good trade-off between simplicity and precision. Code could be easily changed to work with OpenPose though.

Tetsujinfr commented 3 years ago

@Lixz123456 There is a mention in the readme that one can use openpose params (dataset flag), so it is already possible subject to some additional efforts. I think openpose would provide more stable pose estimates for sequences of images from video (same scene), at least this version of it.

I think you can close this "issue" if you agree.

Tetsujinfr commented 3 years ago

Effectively when running Expose with the default RCNN pose estimation on a video sequence the pose estimates are a bit unstable. But the most painful part is the ton of false positive detections, which have to be cleaned up manually.

OpenPose would likely not generate so many false positives, and the version I mentioned in my previous post has good temporal stability as you can see here

Bobby-youngking commented 2 years ago

Hi, keypointrcnn_resnet50_fpn is used for body detection. I don't know why author decided to use this, but probably is a good trade-off between simplicity and precision. Code could be easily changed to work with OpenPose though.

May I ask how to keypointrcnn_resnet50_fpn by my own?