search

varunvprabhu / simple_pose_hourglass

Pose Estimation with Hourglass Model
0 stars 0 forks source link

data convertsion #2

Open siva-wellnesys opened 3 years ago

siva-wellnesys commented 3 years ago

hi i have couple of doubts regarding to training. 1)how did you converted mpii data into image_mpii_data.hdf5 format. 2) in mpii there is point name pelvis , which is come on the navel of the body, but in your model output that point is coming at root(i.e in between the hips) same as neck point also, here in mpii we dont have neck point, it has spine which comes at heart position, but in your case it is coming in between the shoulders. how did you modified this data? 3) I saw your text file data folder you are not using the visibility variable(0-visible,1-not visible,2-occlusion) from the original mpii data.instead of that you are using -1. how did you trained the data without visibility variable? 4) from the 3rd point if you wont use visibility then all point will appear irrespective of the confidence in different positions of image.how did you handled this situation?

varunvprabhu commented 3 years ago

Hi Siva,

While I did this a while back, I'll do my best to remember all the details.

  1. I added a new jupyter notebook in the notebooks folder that should help with understanding the dataset conversion of both the pose data and images to hdf5 format.
  2. The skeleton image below should show the assumptions I made for the point annotations. skeleton
  3. I assumed all joints were visible unless they had the value of -1. If they had the value of -1, I simply ignored it during the rgb image creation.
  4. See point 3.
aliansgp commented 8 months ago

Hello, thank you for sharing the notebooks on GitHub. I'm having trouble understanding how the 'image_mpii_data.hdf5' file is created. I've created the pose_raw file, but I couldn't find any code or help to create the 'image_mpii_data.hdf5' file. Can you please update this section or share the file with me?