search

midusi / LSA-T

https://midusi.github.io/LSA-T/
MIT License
13 stars 0 forks source link

Why was Alphapose chosen? #7

Open yayayru opened 1 year ago

yayayru commented 1 year ago

Why was Alphapose chosen? What about Mediapipe holistic or OpenPose?

pedroodb commented 1 year ago

Alphapose was chosen as it allowed us to use the Halpe dataset format, which we found useful for sign language representation, and also for performance reasons as it can run on 8GB of RAM.

Probably Mediapipe and Openpose could work as well, but we didn't find any particular advantage.

yayayru commented 1 year ago

What is the particular special advantage of Halpe dataset format for sign language representation if you could answer? Would you be able to show an example of sign language representation?

Halpe dataset format, which we found useful for sign language representation

facundoq commented 1 year ago

Err.. it has more finger keypoints than other formats/models. Also Alphapose was (is?) significantly easier to work with since it is based on pytorch.

Can I ask why the interest in such a specific detail? Also, you can always run openpose on the videos...

El jue, 23 feb. 2023 23:11, Alexey Prikhodko @.***> escribió:

What is the particular special advantage of Halpe dataset format for sign language representation if you could answer? Would you be able to show an example of sign language representation?

Halpe dataset format, which we found useful for sign language representation

— Reply to this email directly, view it on GitHub https://github.com/midusi/LSA-T/issues/7#issuecomment-1442697185, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAMA6B6JZCU6TZEFQZ5OPBLWZAKDNANCNFSM6AAAAAAVCU5SL4 . You are receiving this because you are subscribed to this thread.Message ID: @.***>

yayayru commented 1 year ago

Yes, you can ask. The main problem with tracked skeletal fingers for sign language recognition task is that I'm interested in a specific detail. I tried OpenPose before, but there is a devops problem with python. Now I use Mediapipe Hands and Pose, Holistic, it's easier to do something there and it's convenient to analyze the finger skeletal model, see my video example.

I know more about Mediapipe than Alphapose. Alphapose and Mediapipe hands are the same number - 21 finger key points. I'm asking you, do you use a 2D or 3D finger skeletal model from Alphapose? Any other specific detail?:)

Err.. it has more finger keypoints than other formats/models. Also Alphapose was (is?) significantly easier to work with since it is based on pytorch. Can I ask why the interest in such a specific detail? Also, you can always run openpose on the videos...