Use python tensorflow Keras to build CNN and LSTM model, use them to predict static and dynamic hand gesture.
use file change_img_to_YCrCb.py to change pictures into black-and-white pictures, As it is more simple to identify through black-and-white picture.
Using CNN to identify static hand gesture
- neural network structure:
Get test accuracy up to 0.75.
- testing result:
- prediction compare to real label:
Because of using black-and-white picture, the prediction quality isn't good for some similar pictures (see example below):
Using LSTM to identify dynamic hand gesture.
- neural network structure:
Put 10 pictures in a group and label them. Add "swapping left" and "swapping right" gesture (see pictures below), labeled as 11 and 12.
- swapping left:
- swapping right:
Get test accuracy up to 0.6538.
- testing result:
- prediction compare to real label:
LSTM model can predict dynamic gesture well (label 11 and 12), but would made more error on identifying static gesture.