search

rayat137 / Pose_3D

Exploiting temporal information for 3D pose estimation
82 stars 10 forks source link

generation of preds_fed.h5 file #13

Closed ShaminiKoravuna closed 5 years ago

ShaminiKoravuna commented 5 years ago

Hello Rayat!

I have a few questions related to the generation of preds_fed.h5 file

  1. Did you make any changes in the stacked_hourglass network model for generating the 2D pose estimation output?
  2. I'm unable to generate the predictions using the pytorch model. Do you have any tensorflow version of it?
  3. If I have the files for each image in json format how can I convert them to .h5 format?

Thanks a lot for time & Support!

rayat137 commented 5 years ago

Hi Shamini,

Thanks for your interest in our work. Here are the responses to your queries:

  1. No I don't remember making any changes to the structure of stacked hourglass network. Although I did fine-tune on H3.6M dataset.
  2. I used the torch-lua code that the authors provided to generate the 2D predictions. PyTorch wasn't there at that time.
  3. So .h5 format saves everything as a dictionary of matrices. So you first need to read the Json file, and store the data as a numpy tensor. Then when writing to the .h5 file give a name to that matrix.

Hope this answers your concerns. Feel free to contact me if you have any questions.

ShaminiKoravuna commented 5 years ago

Thanks a lot for the quick response. I'll check with that and get back to you if I still have any questions.