A simple app that recognises numbers from hand gestures using a live webcam feed.
This app was trained with a custom dataset built using the video_stream.py file which contains video capture features using OpenCV.
Navigate to the Jupyter Notebook first to get a full rundown of how the system works, create the correct directories to build your own dataset, and a full explanation of how the system is trained.
First, ensure you have created the proper folders to store the data using code from the top of the Jupyter Notebook.
To capture data:
python3 video_stream.py --cat [cat] --n_photos [n_photos]
Specify which category you want to capture data for (1 to 6, 6 being no digits shown) and the number of photos you want to take
Once you have captured the data, return to the Jupyter notebook to split the data out into training, validation and testing data.
To train a new model:
python3 train.py --data-dir images --epochs [epochs]
--data-dir specificies the directory of the training data, and --epochs specifies the number of epochs you want to train your model for. The training function will automatically save a checkpoint after it completes so you can leave it training without needing to worry about the progress.
python3 video_stream.py -i [-t]
-i sets the program to inference mode, and the optional [-t] flag can be selected if you would like to the see the label of the top predicted category as opposed to the output probability for every label.