Question : FPS rate

[dhgokul]

@mganeko How much time it would take to process live video frames @ 60 FPS ?

[mbotsu]

Hi.
If it is a high-speed device it can process in about 0.2 seconds.
That is, it can be processed with 12 sec = 0.2 sec * 60 frame.
Thanks.

[dhgokul]

Thanks mbotsu !

[dhgokul]

@mganeko is there any way to speed up detection process ?

[chen-xi-wang]

Yes, same question as @dhgokul. I am considering that whether this project could use Metal performance shaders to do the part of Matrix Multiplication to improve the process speed. if you guys have any way to speed up it, please to let me know. Thanks @mganeko and @dhgokul !

[mbotsu]

Hi.

Much of the cost is the structure of the machine learning model and the problem of the PAF algorithm.
There is PoseNet as a new approach.
However, this also has a problem of accuracy.
Please try nfocom-tpo / PoseNet-CoreML if you like.
Thanks.

[chen-xi-wang]

Ok. Thanks a lot!

[chen-xi-wang]

Hi mbotsu, I'm still working on this project, that is awesome. but some points of the code I really confused. Could you please provide a development document or some explanation like flow chart? That would be very helpful! Thanks!

[NarutoYq]

Estimator.swift's estimate method uses a Matrix structure, which is a waste of time. Especially during initialization, the matrix data will be copied internally, which can be replaced with an Array and accessed through a pointer. After the last optimization, the same picture needs 200ms before optimization, and only 50ms after optimization。

[mbotsu]

Hi.

I haven't tried it yet, but if you want a faster implementation I recommend this code.
Apache License 2.0

https://github.com/opencv/open_model_zoo/tree/master/demos/human_pose_estimation_demo

and CoreML Model Convert Log.
https://gist.github.com/mbotsu/8c712093884cb64f9477881cb182dbc2

[NarutoYq]

Thank you