LSTM for Action Recognition
Description
This script takes the 3D skeleton as input and trains a 3-layer LSTM. Two models of LSTMs are defined in the model.py script (You can use any one of them). For demo - the location of pre-processed 3D skeleton files are mentioned in the lstm_train.sh script. You can change this location for processing it on your dataset. For other dataset, you also need to change the dataloaders.
REQUIRED PACKAGES AND DEPENDENCIES
- python 3.6.8
- Tensorflow 1.13.0 (GPU compatible)
- keras 2.3.1
- Cuda 10.0
- CuDNN 7.4
Execution
Example- sh lstm_train.sh
Input parameters are provided in options.py By default the parameters are defined for Toyota Smarthome The skeletons are LCRNet output files transformed into numpy arrays.
The script will generate a weight directory in the name of the experiment, where the models will be saved after every epoch. It will also generate a csv file with the training details. The best model should be used for testing using the evaluation_model.py script.
Enjoy AR with LSTM!!!
Reference
[1] S. Das, M. Koperski, F. Bremond and G. Francesca. Deep-Temporal LSTM for Daily Living Action Recognition. In Proceedings of the 14th IEEE International Conference on Advanced Video and Signal-Based Surveillance, AVSS 2018, in Auckland, New Zealand, 27-30 November 2018.