I was looking at the sample in: Riconoscimento dell'azione con una CNN 3D gonfiata | TensorFlow Hub
and works just fine but how to run this against a life feed? is it possible? and how to train my own model if required?
I was looking at the sample in: Riconoscimento dell'azione con una CNN 3D gonfiata | TensorFlow Hub
and works just fine but how to run this against a life feed? is it possible? and how to train my own model if required?
For streaming I suggest you to take a look at the streaming models in:
https://tfhub.dev/google/collections/movinet/
You can finetune these on your data or train from scratch.
Thank you for answering i did review that one but i am a bit confuse, i would like to test this against a webcam on life feed. i dont see how the code would work for this, as i believe this will read the hole video but in live data there is no end.
You can pass a stream chunk as you can see in the example at:
https://tfhub.dev/tensorflow/movinet/a5/stream/kinetics-600/classification/2
You need to access to the camera with your code (Opencv, TFIO, Video4Linux etc…)
Instead If you want to run this on Android you need to use TF lite and write your own demo/example.
You can also try to use Mediapipe if you like:
https://blog.gofynd.com/mediapipe-with-custom-tflite-model-d3ea0427b3c1
Thank you very much, this TFHUB Is new to me but this seems to be the solution.
I really appreciate your time, thank you.