[Japanese/English]
Note
I created a repository of model zoo for keypoint classification.
→ Kazuhito00/hand-keypoint-classification-model-zoo
Estimate hand pose using MediaPipe(Python version).
This is a sample program that recognizes hand signs and finger gestures with a simple MLP using the detected key points.
This repository contains the following contents.
- Sample program
- Hand sign recognition model(TFLite)
- Finger gesture recognition model(TFLite)
- Learning data for hand sign recognition and notebook for learning
- Learning data for finger gesture recognition and notebook for learning
- mediapipe 0.8.1
- OpenCV 3.4.2 or Later
- Tensorflow 2.3.0 or Later
tf-nightly 2.5.0.dev or later (Only when creating a TFLite for an LSTM model) - scikit-learn 0.23.2 or Later (Only if you want to display the confusion matrix)
- matplotlib 3.3.2 or Later (Only if you want to display the confusion matrix)
Here's how to run the demo using your webcam.
python app.pyHere is how to run the demo using Docker and a webcam.
docker build -t hand_gesture .
xhost +local: && \
docker run --rm -it \
--device /dev/video0:/dev/video0 \
-v `pwd`:/home/user/workdir \
-v /tmp/.X11-unix/:/tmp/.X11-unix:rw \
-e DISPLAY=$DISPLAY \
hand_gesture:latest
python app.pyThe following options can be specified when running the demo.
- --device
Specifying the camera device number (Default：0) - --width
Width at the time of camera capture (Default：960) - --height
Height at the time of camera capture (Default：540) - --use_static_image_mode
Whether to use static_image_mode option for MediaPipe inference (Default：Unspecified) - --min_detection_confidence
Detection confidence threshold (Default：0.5) - --min_tracking_confidence
Tracking confidence threshold (Default：0.5)
│ app.py
│ keypoint_classification.ipynb
│ point_history_classification.ipynb
│
├─model
│ ├─keypoint_classifier
│ │ │ keypoint.csv
│ │ │ keypoint_classifier.hdf5
│ │ │ keypoint_classifier.py
│ │ │ keypoint_classifier.tflite
│ │ └─ keypoint_classifier_label.csv
│ │
│ └─point_history_classifier
│ │ point_history.csv
│ │ point_history_classifier.hdf5
│ │ point_history_classifier.py
│ │ point_history_classifier.tflite
│ └─ point_history_classifier_label.csv
│
└─utils
└─cvfpscalc.py
This is a sample program for inference.
In addition, learning data (key points) for hand sign recognition,
You can also collect training data (index finger coordinate history) for finger gesture recognition.
This is a model training script for hand sign recognition.
This is a model training script for finger gesture recognition.
This directory stores files related to hand sign recognition.
The following files are stored.
- Training data(keypoint.csv)
- Trained model(keypoint_classifier.tflite)
- Label data(keypoint_classifier_label.csv)
- Inference module(keypoint_classifier.py)
This directory stores files related to finger gesture recognition.
The following files are stored.
- Training data(point_history.csv)
- Trained model(point_history_classifier.tflite)
- Label data(point_history_classifier_label.csv)
- Inference module(point_history_classifier.py)
This is a module for FPS measurement.
Hand sign recognition and finger gesture recognition can add and change training data and retrain the model.
Press "k" to enter the mode to save key points（displayed as 「MODE:Logging Key Point」）
If you press "0" to "9", the key points will be added to "model/keypoint_classifier/keypoint.csv" as shown below.
1st column: Pressed number (used as class ID), 2nd and subsequent columns: Key point coordinates
The key point coordinates are the ones that have undergone the following preprocessing up to ④.
In the initial state, three types of learning data are included: open hand (class ID: 0), close hand (class ID: 1), and pointing (class ID: 2).
If necessary, add 3 or later, or delete the existing data of csv to prepare the training data.
Open "keypoint_classification.ipynb" in Jupyter Notebook and execute from top to bottom.
To change the number of training data classes, change the value of "NUM_CLASSES = 3"
and modify the label of "model/keypoint_classifier/keypoint_classifier_label.csv" as appropriate.
The image of the model prepared in "keypoint_classification.ipynb" is as follows.
Press "h" to enter the mode to save the history of fingertip coordinates (displayed as "MODE:Logging Point History").
If you press "0" to "9", the key points will be added to "model/point_history_classifier/point_history.csv" as shown below.
1st column: Pressed number (used as class ID), 2nd and subsequent columns: Coordinate history
The key point coordinates are the ones that have undergone the following preprocessing up to ④.
In the initial state, 4 types of learning data are included: stationary (class ID: 0), clockwise (class ID: 1), counterclockwise (class ID: 2), and moving (class ID: 4).
If necessary, add 5 or later, or delete the existing data of csv to prepare the training data.
Open "point_history_classification.ipynb" in Jupyter Notebook and execute from top to bottom.
To change the number of training data classes, change the value of "NUM_CLASSES = 4" and
modify the label of "model/point_history_classifier/point_history_classifier_label.csv" as appropriate.
The image of the model prepared in "point_history_classification.ipynb" is as follows.
The model using "LSTM" is as follows.
Please change "use_lstm = False" to "True" when using (tf-nightly required (as of 2020/12/16))
Here are some application examples.
- Control DJI Tello drone with Hand gestures
- Classifying American Sign Language Alphabets on the OAK-D
Kazuhito Takahashi(https://twitter.com/KzhtTkhs)
hand-gesture-recognition-using-mediapipe is under Apache v2 license.