National Action Council for Minorities in Engineering(NACME) Google Applied Machine Learning Intensive (AMLI) at the PARTICIPATING_UNIVERSITY
Developed by:
- member1 -
Ruben Priest_UARK - member2 -
Donovan Saat_UARK - member3 -
Daniel Overton_UARK - member4 -
Trenton Harris_UARK
Scene hand detection for real world images.
)
To detect hand gestures, we first have to detect the hand position in space. This pre-trained network is able to extract hands out of a 2D RGB image, by using the YOLOv3 neural network.
There are already existing models available, mainly for MobileNetSSD networks. The goal of this model is to support a wider range of images and a more stable detector (hopefully 🙈).
The first version of this network has been trained on the CMU Hand DB dataset, which is free to access and download. Because the results were ok, but not satisfying, I used it to pre annotate more images and manually then corrected the pre-annotations.
Because Handtracking by Victor Dibia is using the Egohands dataset, I tried to include it into the training-set as well.
In the end, the training set consists of the CMU Hand DB, the Egohands dataset and my own trained images (mainly from marathon runners), called cross-hands.
The training took about 10 hours on a single NVIDIA 1080TI and was performed with the YOLOv3 default architecture. I also trained the slim version of it called YOLOv3-tiny.
Precision: 0.89 Recall: 0.85 F1-Score: 0.87 IoU: 69.8
Precision: 0.76 Recall: 0.69 F1-Score: 0.72 IoU: 53.67
The tiny version of YOLO has been improved by the partial residual networks paper. Because of that I trained YOLO-Tiny-PRN and share the results here too. It is interesting to see that the Yolov3-Tiny-PRN performance comes close to the original Yolov3!
Precision: 0.89 Recall: 0.79 F1-Score: 0.83 IoU: 68.47
With the recent version of YOLOv4 it was interesting to see how good it performs against it's predecessor. Same precision, but better recall and IoU.
Precision: 0.89 Recall: 0.89 F1-Score: 0.89 IoU: 91.48
I could not test the model on the same dataset as for example the Egohands dataset, because I mixed the training and testing samples together and created my own test-dataset out of it.
As soon as I have time, I will publish a comparison of my trained data vs. for example Handtracking.
The models have been trained on an image size 416x416. It is also possible to inference it with a lower model size to increase the speed. A good performance / accuracy mix on CPUs has been discovered by using an image size of 256x256.
The model itself is fully compatible with the opencv dnn module and just ready to use.
To run the demo, please first install all the dependencies (requirements.txt) into a virtual environment and download the model and weights into the model folder (or run the shell script).
# mac / linux
cd models && sh ./download-models.sh
# windows
cd models && powershell .\download-models.ps1
Then run the following command to start a webcam detector with YOLOv3:
# with python 3
python demo_webcam.pyOr this one to run a webcam detrector with YOLOv3 tiny:
# with python 3
python demo_webcam.py -n tinyFor YOLOv3-Tiny-PRN use the following command:
# with python 3
python demo_webcam.py -n prnFor YOLOv4-Tiny use the following command:
# with python 3
python demo_webcam.py -n v4-tiny- YOLOv3 Cross-Dataset
- YOLOv3-tiny Cross-Hands
- YOLOv3-tiny-prn Cross-Hands
- YOLOv4-Tiny Cross-Hands
If you are interested in the CMU Hand DB results, please check the release section.
Trained by cansik, datasets are described in the readme and fall under the terms and conditions of their owners.
All the demo images have been downloaded from unsplash.com:
Tim Marshall, Zachary Nelson, John Torcasio, Andy Falconer, Sherise, Alexis Brown
This README.md is credited to the github user "cansik" for the hand detection model using YOLOv3
Please reach out to them for any questions via github