Hand tracking application developed in Pytorch.
Hand tracking application developed for an university project. It used a RCNN (Regions with CNN) to track hands inside the image and a CNN (Convolutional Neural Network) to evaluate the number of the fingers in the hand.
The RCNN training script is based on the script at tutorial in the Pytorch site. I use a pretrained RCNN on COCO dataset and with fine tuning I teach the network how to recognize hands. The original tutorial used a Mask-RCNN that also create a mask of the shape of the things you search. Since I was only interested in a square box aruond hand I modify the script to use a simpler RCNN.
To test the track in real time you also need OpenCV.
A simple video demo is available at https://www.youtube.com/watch?v=1nmR-dRdeos.
I try two different dataset to train the network. The firs is the pretty famous EgoHands Dataset (The 1.3 Gb version) and the other is a my personal (costum) dataset that I create. The results are similar and you could choose which one to use. The trained model provided in the link is trained on my personal dataset.
To create your own dataset used the file inside the RCNN (my dataset) folder. The files are well commented but basically you need to:
- Execute the dataset_creator.py (you need OpenCV for this). This script create a windows with the image capture by your camera and n boxes (n set by you). You position the hand inside the boxes and after x seconds (x decide by you) the image and the boxes position are saved.
- Check the dataset with dataset_check.py (OPTIONAL).
- Train the network with train_RCNN.py .
The MyDataset.py file contain the Dataset class used during training. This class is an extension of the Pytorch Dataset class and it will used in combination with the Dataloader (more info and example here, here, here and here).
The support_function.py contain function used during the dataset creation and the dataset loading.
NB: The train need torchvision. All the file are inside the vision folder and that folder must be in the same folder of train_RCNN.py.
To train the model with the EgoHands dataset download the dataset from the link (The 1.3 Gb version). Extract it and put the folders inside the _LABELLED_SAMPLES in tha path that you will use to contain the file for the training. After that execute the file train_RCNN.py .
NB: The train need torchvision. All the file are inside the vision folder and that folder must be in the same folder of train_RCNN.py.
The CNN is a simple network trained on a costum dataset. The dataset is created with the file dataset_creation.py in a similar way to the costum dataset for the RCNN. I use an Adadelta optimizer with a Cross Entropy loss function. For some reason sometimes the training succed and sometimes remain stuck and don't learn after hundreds of epoch. In the last case stop and restart the training.
I use 140 x 140 square as size of the image but you could choose the size that you prefer.
The trained network provided with the code doesn't work very well since I train its on 200 example. Also I suggest you to use always the same hand to improve the accurancy. A trick later used in the the final script is to mirror one of the two hand find by RCNN so the classifier will only receive one type of hand and give similar results for both hand.
The hand_detector_V2.py contain the script that track the hand(s). It use OpenCV to capture the camera image and feed that image to the RCNN. With the RCNN I find the hand, take only that portion of the image, reshape in a 140 x 140 square and feed them into the fingers counter.
NB: You need a GPU with CUDA to obtain a (more or less) smoot frame rate. I use a GTX 2070 but probably even less powerfull GPU work. To improve the FPS I use a little trick: instead of feed every frame into the RCNN I do it every 3 frame. In this way the code works smoothly granted that you had a GPU capable of executing the code required by the RCNN in relative small time.
- Add video explanation.
- Modify the costum dataset to label differently left hand and right and.
- Retrain the RCNN to allow recognition of left and right hand.
If you ever use this code please cite me 🙏 ❤️
If you need to contact my you could write me at alberto.zancanaro.1@gmail.com or alberto.zancanaro.1@studenti.unipd.it .