RNNforGestures

This repo contains materials, source code and usage description of a script, which trains and evaluates Recurrent Neural Networks (RNNs) applied for hand gesture recognition.

Overview of readme:

• Description
• Usage
• References

Description

The hand gesture recognition with sensors of low complexity and of low power demands is of interest especially in the context of mobile devices. The RNNrandHyp.py script available in this repo is dedicated to optimize the hyperparameters of a Recurrent Neural Network in order to recognize a set of 27 gestures with the highest possible accuracy. The input data comes from the optical linear sensor described in [1-2].

Some of the analysis of the results are presented in the notebook.
More detailed description of the sensor, results and discussion are presented in our article from the IEEE Sensors Journal [3].

Usage

The RNNrandHyp.py python script needs following packages to be installed:

sklearn, tensorflow, numpy, pandas, Enum, enum34, datetime, sys

In order to run the script from console type:

python RNNrandHyp.py my_option my_gpu_memo my_hml

where with my_option equal to:

• 0 - a single run on raw dataset, with predefined values of hyperparameters is evaluated,
• 1 - raw dataset is selected, hyperparameters are sampled,
• 2 - features dataset is selected, hyperparameters are sampled,
• 3 - HLfeatures dataset is selected, hyperparameters are sampled,
  

with my_gpu_memo in the range of 0 to 100:

• a given percentage of GPU card memory is allocated for the script. In practice, less than 4GB is enough in this case. Therefore, if one has a 4GB card the my_gpu_memo could be set to 80 or 100, whereas for 16GB 20 or 25,
  

with my_hml:

• a number telling the script how many loops with random search trials should be executed. If you wish to execute 100 trials and have 16GB card, you can run 4 scripts per 25 trials and my_gpu_memo=25.

Example:

In order to evaluate 16 trials on raw dataset, with randomly sampled hyperparameters on 0th GPU card, allocating 25% of its memory run:

CUDA_VISIBLE_DEVICES=0 python RNNrandHyp.py 1 25 16

---

References

[1] "Analysis of Properties of an Active Linear Gesture Sensor" K. Czuszynski, J. Ruminski, J. Wtorek
[2] "Pose classification in the gesture recognition using the linear optical sensor" K. Czuszynski, J. Ruminski, J. Wtorek
[3] "Gesture Recognition with the Linear Optical Sensor and Recurrent Neural Networks" K. Czuszynski, J. Ruminski, A. Kwasniewska