search

facebookresearch / VideoPose3D

Efficient 3D human pose estimation in video using 2D keypoint trajectories
Other
3.72k stars 755 forks source link

Recommended machine and Google Colab #43

Open Uiuran opened 5 years ago

Uiuran commented 5 years ago

Hi,

I would like to know the minimal settings to run this model, maybe with a little more of training, and if it is possible to do it in Google Colab.

I will be doing this for a martial-arts group of people (Kung Fu - Aikido). However my machine sux ...

Thank you for your attention.

dariopavllo commented 5 years ago

I'm not familiar with Google Colab, but I see no reason why it shouldn't work.

Regarding the minimal settings, for inference any GPU will do fine (even a laptop GPU). For training (especially for the models with a large receptive field), I would suggest at least a Pascal GPU. It also depends on whether you want to extract 2D keypoints from videos (e.g. using Detectron). If so, a Pascal GPU would be a safe bet.

Justinemmerich commented 5 years ago

I have created a Google Colab Jupyter Notebook for this. You can get it here: Google Colab Notebook - Facebook VideoPose3D Inference in the wild

Go to - https://colab.research.google.com/ click on Github Tab add the ipynb link

Justinemmerich commented 5 years ago

@Uiuran hope this helps!

Justinemmerich commented 5 years ago

@dariopavllo this should solve it! maybe I could request a pull next time.

aowais2 commented 4 years ago

@Justinemmerich Just wanted to thank you for your contribution. The Google Colab works like a charm

immkapoor commented 4 years ago

@Justinemmerich Hey, I tried running the colab notebook, but I got an error while computing 3D joints as: ImplementedError: It is not currently possible to manually set the aspect on 3D axes. Do you know anything that might have caused this error?

slava-smirnov commented 4 years ago

@immkapoor it has nothing to do with collab itself, new matplotlib is messing with setting axes as it's heavily discussed on their issues. however your issue is solved with commenting #ax.set_aspect('equal') on L91 at visualisation.py

immkapoor commented 4 years ago

@slava-smirnov Hi, going back to older version of matplotlib resolved the issue for me.