This repository contains the original implementation of the paper Neverova et al., "ModDrop: adaptive multi-modal gesture recognition", IEEE TPAMI, 2016 (https://arxiv.org/abs/1501.00102).
Original Theano implementation by Natalia Neverova natalia.neverova@gmail.com
Re-implementation of DeepGestures using Lasagne by Dhanesh Ramachandram dhaneshr@gmail.com
The code was tested with Theano 0.82 and Lasagne 0.2dev1 and Anaconda Python on Ubuntu 14.04 running CuDNN version 4.0
If you use this code in your research, please cite the original manuscript:
@artucle{moddrop,
author={Natalia Neverova and Christian Wolf and Graham Taylor and Florian Nebout},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
title={ModDrop: Adaptive Multi-Modal Gesture Recognition},
year={2016},
volume={38},
number={8},
pages={1692-1706}
}
- Bleeding-edge version of Theano Install using :
#!python
pip install --upgrade --no-deps git+git://github.com/Theano/Theano.git
- Bleeding edge version of Lasagne Install using :
#!python
pip install --upgrade --no-deps https://github.com/Lasagne/Lasagne/archive/master.zip
- CuDNN Download CuDNN v4 from NVIDIA website. Extract the library into a folder on your home directory. Then add the following lines in ~/.bash_profile
#!bash
export LD_LIBRARY_PATH=/path/to/cuda/lib64:$LD_LIBRARY_PATH
export CPATH=/path/to/cuda/include:$CPATH
export LIBRARY_PATH=/path/to/cuda:$LIBRARY_PATH
-
training_script.py is the main script to initiate training. In this file, modify the path to the Montalbano dataset accordingly. It should point toward a preprocessed version of the Montalbano dataset which can be found at http://hdl.handle.net/10864/9VNIK. We are working on cleaning up and releasing a set of scripts to provide the full pre-processing routines to convert the original data to our format. The original data is available at http://sunai.uoc.edu/chalearnLAP/ (Track 3: Gesture recognition).
-
initiate the training by calling
#!python
python training_script.py modality step
Replace modality with 'skeleton' for mocap, 'audio' for audio, 'videoFeat' for video,depth pretraining, 'video' for video-depth fusion/pre training and finally 'multimodal' for all modalities. skeleton and audio can be trained independently, but videoFeat must be trained, before video.
All other modalities must be pre-trained before multimodal.
'step' can be 2 or 4, but in practice, 4 gives best results.
Also, required for testing phase, a classifier called motionDetector must be called with step = 1 using the training script.
In the same folder as your source-code, there must be several subfolders created.
- subfolder 'filters' - used to store saved models. There is a file called skeleton_stats which must be present in the folder.
- subfolder 'ground_truth' which contains the ground_truth csv files used for testing.