18 months ago I admitted defeat trying to teach Dave to understand American Sign Language with his “naked Pi4”.
Of course I was trying to jump straight to the end of the “train your robot” process that
- starts with data…lots and lots of data,
- continues with processing that lots of data to “build a model”, and
- ends with using the model in a specialized neural network processor to recognize something “familiar” in new visual frames.
Saw an article on the datasets for several Google sponsored competitions for machine learning:
- American Sign Language gestures and
- American Sign Language finger spelling.
The input data for the ASL competitions consists of “only” 61 million samples in 250GB.
I have resisted building a machine learning environment so am limited to using available models.
Hopefully these Google sponsored ASL competitions will result in TFlite models to run on Dave’s Oak-D-W Neural Net processor to give him a sophisticated human interface input modality.