Air-writing recognition from an image stream input

This repository contains an example code used for air-writing recognition from an image input viewpoint. That means there is no previous hand-localization stage, but the class of the performed gesture is predicted from the whole video sequence.

Installation on Windows 10

1. Install Anaconda for Python 3.6: Anaconda.
2. Download this repository.
3. (If not installed) Install CUDA Toolkit, Nvidia drivers, and library cuDNN for GPU support in Tensorflow. More instructions in section Requirements to run TensorFlow with GPU support from Tensorflow Installation Guide.
4. Install the conda environment and required packages: "conda env create -f tensorflow.yml".
5. Download and install "graphviz-2.38.msi" from https://graphviz.gitlab.io/_pages/Download/Download_windows.html.
6. Add the graphviz bin folder to the PATH system environment variable (Example: "C:/Program Files (x86)/Graphviz2.38/bin/")
7. Create the subfolder "models".
8. Develope dataset from the link Leap Motion writing acquisition to the subfolder "input". The final dataset will have with the following folder structure:

gesture_1/
  repetition_1/
    frame_000000.png
    frame_000001.png
    ...
    frame_000999.png
    
  repetition_2/
  ...
  repetition_7/
  
gesture_2/
  ...
  
gesture_N/

where repetition_N is a sample folder and class_N is a writing gesture type.

Running the code on Windows 10

Prediction from a trained model

Execute ./windows/testme.bat Alternatively:

1. Run a Anaconda prompt.
2. Activate the conda environment with the command "activate tensorflow".
3. Execute:

python ../test.py --experiment_rootdir=../models ^
--weights_fname=../models/test_4/weights_015.h5 ^
--img_mode=rgb

Note1 : Depending on your installation, you will need to write python3 or just python to run the code.

Training from scratch

Execute ./windows/trainme.bat Alternatively:

1. Run a Anaconda prompt.
2. Activate the conda environment with the command "activate tensorflow".
3. Execute:

python train.py --experiment_rootdir=./models/test_1 ^
--img_mode=rgb 

See more flas in common_flags.py to set batch size, number of epochs, dataset directories, etc.

Training from an existing model (fine-tuning)

1. Run a Anaconda prompt.
2. Activate the conda environment with the command "activate tensorflow".
3. Execute:

python train.py --restore_model=True --experiment_rootdir=./models/test_1 ^
--weights_fname=models/weights_015.h5 ^
--img_mode=rgb 

where the pre-trained model called m ./models/test_1 must be in the directory you indicate in --experiment_rootdir.