OpenHands | WIP open source hand

OpenHands is a Python library. OpenHands has no bugs, it has no vulnerabilities, it has build file available and it has low support. You can download it from GitHub.

WIP open source hand tracking. I'm currently re-training the main pose model and will add links to the dataset (50k + images) and Tensorflow training code. Sometimes the hand bounding box will continually shrink, you will have to wait for it to re-track the hand when this happens. It's likely due to the image padding when training, I'm working on this atm. The demo only tracks one hand, if you mask this hand with a gray box and feed the image to a second detector you will get two hands. Architecture: Detection model: Standard CNN, 256x256x3 input, continuall down samples, dense layer on top for hand position and confidence. Pose: Uses 64x64x21 heatmap of points and vector of (21, 2) to emulate the previous frames predicted frame (for stability/jitter). Standard CNN, input size: 128x128x3 downsamples to 8x8 then uses Conv2DTranspose to upsample to the input resolution. Uses argmax to predict points on heatmap, which are then fed to a dense top layer, which also predicts the 3d co-ordinates. I also have a stacked hour glass model which is more accurate (preserves more spatial information through skip connections) but is around 2-3x slower depending on your hardware.