zh-plus/video-to-pose3D

Tensorflow OpenPose implementation

sebo361 opened this issue · 4 comments

sebo361 commented

Hi @zh-plus, i am happy to help you with finalizing the openpose detector. As I am using Anaconda I have problems with installing openpose as Anaconda includes a Protobuf version that is incompatible with Caffe. Therefore I thought it would be easier to use this tf openpose implementation. Finally I could write a conda install file to easily set up the system. What do you think?

zh-plus commented

Since the accuracy of openpose is not better than AlphaPose, I'm not planning to add it into project. Instead, I'd like to wait for a new SOTA 2d pose estimation method later this year.

sebo361 commented

Thanks for your comment @zh-plus!

So I assume you the duc_se.pth model is the one with ResNet152 backbone (most accurate regarding this table)? I looked into Alphapose repo and found a --flip flag which might improve accuracy but its not used in this repo right?

I am looking for a configuration to get the most accurate single person 2D keypoint detection available (no matter how long inference take), any idea on how to improve / add another model for single person 2D?

zh-plus commented
  1. The table is added 3 months ago after I found the duc_se.pth model. I can't remember where I found it. I believe there'll be an improvement replacing it with this most accurate model.
  2. flip flag is also newly added here in the new version.
  3. Currently, I'm not catching up with the latest released, fancy papers. Thus I'm afraid that I can not provide any idea on it.
sebo361 commented

Thank you @zh-plus for the information and your input. I will dive deeper into it and let you know / create PR. Good luck with your thesis!