self portrait

Tracking Hands

The last few months I’ve been working on building a small program that can track hands from a webcam feed in real-time. For this, I used Deep Neural Networks, with many thanks to for getting me caught up on the subject. Getting up and running with Convolutional Neural Networks was pretty easy, as was training and getting decent accuracy with a limited training set. I kept a more detailed log in this forum thread. The results below were generated frame-by-frame and then converted to a movie.

Example GIF

Unfortunately, the real problem with this project is turning out to have a more technical (as opposed to theoretical) nature. My goal was to wrap the neural network in a somewhat-portable “tracking server” that broadcasts to other motion-controlled apps on the same network/machine. I have tried OpenCV::DNN and tiny-dnn in an openFrameworks app, but I haven’t found a convenient solution to using the network yet. From what I’ve seen, the research libraries (Keras, PyTorch) are elegant and easy to use, but using the networks from C++ in ‘production’ is a whole different story.