/3d-handtracking-demos

A collection of 3D Hand Pose estimation models

Primary LanguagePythonMIT LicenseMIT

3D Hand Tracking Demos

Biggest names in the tech industry (Microsoft, Facebook, SnapChat) have released commercial products that support 3D hand tracking without any controllers OR markers. This library will allow users to demo multiple existing open-source 3D hand tracking solutions for monocular RGB cameras.

TODO List

In descending order of priority

Hand Detectors

  • Standalone hand detection demos
  • Standardise hand detector outputs by creating a wrapper (possibly shared parent)
  • Crop images based on hand detector output (Might be difficult with BlazePalm output being non-axis-aligned)
  • Figure out the best way to incorporate detector within the architecture so that it's switched on for some models (e.g. Meshformer) and switched off for others e.g. E2E Mediapipe tracking, minimal-hand

Keypoint Estimators

  • Implement Meshformer
  • Implement FreiHand Estimator
  • Implement Minimal Hand

GUI Improvements

  • Allow users to drag-and-drop custom models (impose limitations on model's I/O)
  • Create a 3D representation of the estimated keypoints/mesh

General Stuff

  • Arguments for the architecture/designed patterns used
  • Possible issues with LICENSE with GANerated Hands and MANO mesh

References

Hand Detection Models

  1. https://github.com/aashish2000/hand_tracking
  2. https://github.com/cansik/yolo-hand-detection

Keypoint Estimation Models:

  1. https://google.github.io/mediapipe/solutions/hands