google-ai-edge / mediapipe

Cross-platform, customizable ML solutions for live and streaming media.
https://mediapipe.dev
Apache License 2.0
26.77k stars 5.09k forks source link

BlazePose 3D points? #1099

Closed vpenades closed 4 years ago

vpenades commented 4 years ago

I've been using BlazePose for a while now... and I've noticed the model outputs 4 float values per landmark.

At first I used to think that the first 2 floats where the XY coordinates.... but by plotting the XYZ coordinates using the first three values, I've noticed there's some kind of 3Dish support, at least with the joints that represent the torso and limbs.

Does it mean that BlazePose outputs (limited) 3D landmarks?

BetaPundit commented 4 years ago

In my opinion, the 3rd value is the confidence score of each key point. Not sure though.

mgyong commented 4 years ago

Yes there are 4 float values per landmark (x, y, z, visibility). As of 9 Sept 2020, z value is not accurate and should not be used in production.

vpenades commented 4 years ago

@mgyong That's very good to know.

Although you recommend not to use it for production, I need to know, in a broad way, the general direction the body is pointing at. Specially, to know if the body is looking to the right or to the left, so most probably, with the current accuracy I can work something out.

Thanks a lot, great work!