This is a Pytorch implementation of Deep Soli paper:
"S. Wang, J. Song, J. Lien, I. Poupyrev and O. Hilliges, "Interacting with soli: Exploring fine-grained dynamic gesture recognition in the radio-frequency spectrum", Proc. 29th ACM Annu. Symp. User Interface Softw. Technol., pp. 851-860, 2016."
The original authors github is a bit complicated and uses many different packeges. This project is a simple implementation which allows parsing the radar data and training a CNN-LSTM neural network for hand gesture classification.
Data and train/validation split configuration were taken from paper's author github: https://github.com/simonwsw/deep-soli
After downloading the data, extract it (as defualt) to \dsp folder.
The preprocessing script 'data_pp_tools.load_data()' loads and parse the data. The data is splite roughly 50/50 to train/validation by the same configuration file from the original authors. Each data point is a sequence of variable length of 1024 dimensional vector which are reshaped to a [32,32] 2d Doppler map [range, velocity].
The preprocessing script truncates the data points to equal length of 40 (as the authors mention in their paper). 33 & 23 sequences are droped from the training & validation sets respectivly as they have less than 40 samples.
Basic parameters can bee seen at the bottom of 'train.py' packed by parser.
'train.py' will load and process the data as described above, and train a CNN-LSTM model. Predictions are calculated based only on the last hidden LSTM state. Training is done using Negative log loss using true labels. Run command for training with default configurations:
python train.py
For each run, a folder named by date and time will be created with a log file containing info about the parameters. Losses arrays, accuracy arrays, figures and the trained model are saved there as well.
