[✔ WORKING]
(See the TODO list below for future improvements)
A program to recognize hand pose from an RGB camera.
These instructions will help you setting up the project and understanding how the software is working. You'll see the file structure and what each file does.
See the requirements.txt file or simply run:
pip install -r requirements.txt
.
├── cnn Contains the cnn architecture and the models.
│ └── models Trained models.
├── Examples
├── hand_inference_graph
├── model-checkpoint
├── Poses The poses dataset. Each pose will have its folder.
│ ├── Fist
│ ├── Four
│ ├── Garbage
│ ├── Palm
│ ├── Rock
│ └── Startrek
├── protos
├── Results
└── utils
To run the multithreaded hand pose recognition, simply run:
python HandPose.py
The mediafire link is here: http://www.mediafire.com/file/wt7dc5e9jgnym04/Poses.tar.gz/file Download, and extract the Poses folder that you then place in the root of the Handpose folder.
OR, on Linux, just run:
./download_dataset.sh
this will download, extract the files and remove the archive file.
To add a new pose, launch the AddPose.py script doing:
python AddPose.py
You will then be prompted to make a choice. Type '1' and 'enter'. Now you can enter the name of your pose and validate with 'enter':
Do you want to :
1 - Add new pose
2 - Add examples to existing pose
3 - Add garbarge examples
1
Enter a name for the pose you want to add :
Example
You'll now be prompted to record the pose you want to add.
Please place your hand beforehand facing the
camera, and press any key when ready.
When finished press 'q'.
Place your hand facing the camera, doing the pose you want to save and press enter when ready. You'll see the camera feed. Move your hand slowly across the frame, closer and further from the camera. Try to rotate a bit your pose. Do every movement slowly as you want to create ghosting.
You can record for as long as you want, but remember that camera_fps x seconds_of_recording images will be generated.
See an example below:
Then you want to head to the new pose folder situated in Poses/name_of_your_pose/name_of_your_pose_1 and manually delete images that doesn't show well your hand pose.
You can optionnally bulk rename them once you finished cleaning but note that it's not required.
Once that is done you want to normalize those newly created images. Launch normalize.py with:
python normalize.py
This script will go to the poses folder and make sure every images is the right size. It will skip those that are already 28x28.
You then have to retrain the network. For that, open the file situated in 'cnn/cnn.py' and edit the hyperparameters and the model file name if needed. The saved model will be situated in 'cnn/models/'
You don't have to specifiy the number of classes, it will be infered from the number of directories under 'Poses/'.
Launch the training with:
python cnn/cnn.py
Garbage examples are examples where you face the camera and don't do any special hand pose. You want to show your hand, move them around, but don't do any of your poses. The goal is for the SSD to detect some hands and also some false positives. This will generate images that aren't any pose, they are garbage. We do that because we don't want our CNN to missclasify every time a hand is seen.
Launch the AddPose.py script doing:
python AddPose.py
You will then be prompted to make a choice. Type '3' and 'enter'.
Do you want to :
1 - Add new pose
2 - Add examples to existing pose
3 - Add garbarge examples
3
You'll now be prompted to record the pose you want to add.
Please place your hand beforehand facing the
camera, and press any key when ready.
When finished press 'q'.
Same thing as before, press 'enter' to start the recording and stop with 'q'. Then normalize and relaunch training.
The pipeline of this project consists of 4 steps :
- A frame is grabbed from the camera by a dedicated thread, converted to RGB (from BGR) and put into the input queue
- A worker grabs the frame from the queue and pass it into the SSD. This gives us a bouding box of where the hand(s) is and the corresponding cropped frame.
- This cropped frame of the hand is then passed to the CNN, which give us a class vector output of values between 0 and 1. These values correspond to the probability of the frame to be one of the classes. The worker has finished its job and put: the frame with bouding box drawn on top, the cropped frame and the classes into three different queues.
- The main thread, responsible of showing the results can grab the informations from the queues and display them in three windows.
Input image 28x28x1 (grayscale). Two convolutionnal layers with ReLu activation and kernel size 3, followed by a 2x2 max pooling. Finally a 128 dense layer followed by a softmax layer to give the 6-classes prediction.
Note: This photo represents the original SSD architecture which uses VGG16 as a feature extractor. The one used in this project is using MobileNet instead.
For more information on the SSD, head to the references
With 4 workers, I achieved 25fps on a intel i5-8300H running @4Ghz.
If using Tensorflow 2, replace import tensorflow as tf with import tensorflow.compat.v1 as tf and add tf.disable_v2_behavior() at the beginning of the script.
When trying to add a pose with AddPose.py, if the video is not being written, try to change the codec from XVID to MJPG in a .mp4 container.
Replace:
fourcc = cv2.VideoWriter_fourcc(*'XVID')
With
fourcc = cv2.VideoWriter_fourcc('M','J','P','G')
with '.mp4' as the extension.
- ⌛ Improve hand detection of the SSD
- ✔
Add instructions for garbage pose - ✔
Update file structure - ✔
Generate requirements.txt - ✔
Clean imports - ✔
Add explanations on the pipeline - ✔
Remove garbage as a choice when adding more example to existing pose - ✔
Add SSD architecture - ✔
Add NN architecture - ✔
Understand why multithreading doesn't work on linux - ✔
See if Keras is the right version (Windows and Linux) - ✔
Fix multi-threaded detection - ✔
Add more examples to each gesture - ✔
Add interface to live see inference from network - ✔
Test model - ✔
Tweak training/structure of CNN
- Victor MEUNIER - HandPose - MrEliptik - victormeunier.com - blog.victormeunier.com - victormeunier.dev@gmail.com
Want to support me? Buy me a coffee!
- Victor Dibia - Real-time Hand-Detection using Neural Networks (SSD) on Tensorflow, (2017)
GitHub repository, https://github.com/victordibia/handtracking
This project is licensed under the MIT License - see the LICENSE file for details