This repository holds keras and pytorch implementations of the deep learning model for hand gesture recognition introduced in the article Deep Learning for Hand Gesture Recognition on Skeletal Data from G. Devineau, F. Moutarde, W. Xi and J. Yang.
Notebooks with the model definition in either pytorch or keras are provided on Google Colab. The notebooks include gesture data loading, model creation, and model training. For offline use in jupyter, you can download the notebooks from colab.
| Description | Google Colab |
|---|---|
| 1. Data Loading | https://colab.research.google.com/drive/1ggYG1XRpJ50gVgJqT_uoI257bspNogHj |
| 2. Model: Keras | https://colab.research.google.com/drive/1DA3qlnT-HlN5aTvWbWT-ARDfaQIEZGfH |
| 2. Model: PyTorch | https://colab.research.google.com/drive/1TcfF3sNBOAXkQC5XU7tHMBfm0FYIC0UR |
Additional details are provided below.
A deep learning model, i.e. a neural network (middle), is used to classify hand gestures. The neural network uses a sparse representation of the hand (left). The neural network extracts motion features, using a dedicated temporal feature extractor (right) made of temporal convolutions. These temporal features are finally used to determine the nature of the gesture performed.
Studies on human visual perception of biological motion (Johansson 1973) have shown that humans can recognize human body motion actions, using the motion of the body's (skeletal) joints positions only, as you can see on this youtube video.
Such skeletal (“pose”) representations are lightweight and very sparse compared to image and video representations.
Some sensors directly provide streams of body skeletons or hand skeletons: e.g. Leap Motion, Kinect camera, RealSense camera, or motion capture suits and gloves. It is also possible to extract the pose information from videos using vision-based approaches like OpenPose, AlphaPose, or Google’s Media Pipe at a good frame rate.
Each hand joint typically has 2 or 3 dimensions, to represent its (x,y) or (x,y,z) position in space at a given timestep. A gesture is thus represented by a sequence over time of n_joints (e.g. 22 joints in the image above) joints, or, equivalently by a sequence over time of n_channels (e.g. 66 channels = 22 joints x 3 channels: for x, y and z position of the joint).
The model use such sequences as input.
The key idea of the model is to extract relevant features to classify the gesture, based on the temporal evolution of each channel/signal. (The extracted features will later be used to perform the classification.)
Each channel is processed separately.
For performance, the temporal feature extraction processing is split into three parallel branches that are later merged.
Let’s describe the first one (left). To extract temporal features for each individual 1D channel (e.g. let’s say the channel representing the y position of the wrist), the neural network uses 1D convolutions over time. The network use 3 convolutional and pooling layers to get a better representation.
In order to deal with different time resolutions, this processing branch is actually present twice (left and right) in the temporal feature extraction module, but with different sizes –time resolutions– for the convolution kernel.
A third, pooling-only, branch (middle) is added in order to help the backpropagation during the training.
Finally, for each channel, the outputs computed by the three branches are concatenated into a single output.
Once features have been extracted for each channel, they need to be “merged”.
To that extent, they are all fed into a dense neural network (one hidden layer) which performs the final classification.
The full model (by-channel temporal feature extraction + final MLP) is differentiable and can be trained end-to-end.
Recognizing hand gestures can be useful in many daily real-life situations: writing, drawing, typing, communicating with sign language, cooking, gardening, driving, playing music, playing sport, painting, acting, doing precise surgery, pointing, interacting with one’s environment in augmented reality or virtual reality, for drone control, lights control, sound control, home automation, medicine, nonverbal communication, … the list is almost limitless!
If you only have videos of the gestures, first use a vision-based approach to generate hand pose/skeletal representation of your sequences.
For instance, you can use the CMU's OpenPose demo: https://github.com/CMU-Perceptual-Computing-Lab/openpose or Google's MediaPipe https://github.com/google/mediapipe/blob/master/mediapipe/docs/hand_tracking_mobile_gpu.md
You can also use motion capture gloves, or dedicated sensors (Leap Motion, RealSense camera, ...) that directly provide such sequences.
If you don't have any hand gesture available or if you want to reproduce the results of the research article, you can also download the DHG-14/28 Dataset or the SHREC’17 Track Dataset. To preprocess the two datasets, and load them with a single line of code, please follow the instructions provided here.
The model expects gestures to be tensors of the following shape: (batch_size, duration, n_channels).
For instance, if you want to recognize a batch of 32 gestures of length 100 on a hand skeletal constituted of 22 joints in 3D (x,y,z), i.e. 22x3=66 channels, the shape of the tensor should be : (32, 100, 66).
First change the load_data() and the functions just below it to load your own data (section 2 of the notebook).
Then load it (section 4 of the notebook).
Specify how many gesture classes you have and how many hand channels (not joints) you have (section 4 of the notebook).
You can now train the model on your own model.
-
You should always visualize the metrics (loss, accuracy, ...; both for train/test/validation) of the model during the training. You can use
tensorboardfor that (see notebook for more details). -
If the validation error starts increasing you've likely already been overfitting for a few epochs
-
If you encounter some
pytorcherror, ensure that your labels are in the correct range of values, and that the gesture tensor shape is correct -
If the accuracy curves are too noisy or not smooth enough, consider decreasing the optimizer's learning rate
-
You can use data augmentation to improve the model performance
The notebook will run fine with:
-
python 3 -
pytorch 1.0and above -
Usual pip modules:
numpysklearnscipy
Optional: if you want to monitor the neural network’s metrics during training, you’ll also need tensorboardX and tensorboard (the latter is provided by tensorflow)
If you find this code useful in your research, please consider citing:
@inproceedings{devineau2018deep,
title={Deep learning for hand gesture recognition on skeletal data},
author={Devineau, Guillaume and Moutarde, Fabien and Xi, Wang and Yang, Jie},
booktitle={2018 13th IEEE International Conference on Automatic Face \& Gesture Recognition (FG 2018)},
pages={106--113},
year={2018},
organization={IEEE}
}