Hand Joint Angle Detection Tool
Budget: £10 – £500 GBP
I have collected hundreds of standard photographs of hands in many poses and from many people. Nothing is uniform: skin tones, hand sizes, lighting, backgrounds, and even finger counts may vary. The goal is to turn this dataset into a practical analysis tool that can:
• calculate the flexion/extension angle at every finger joint in each image
• flag when a finger or phalanx is completely absent so that it can be recorded automatically
I expect you will combine classical computer-vision preprocessing with a deep-learning model (PyTorch or TensorFlow are fine) and possibly leverage pose-estimation libraries such as MediaPipe or OpenCV for landmark detection. Accuracy matters more than perfection in lighting or background, so robust data-augmentation and domain-adaptation techniques will be key.
Phase 1
Train and validate a model that ingests a single hand photograph and returns a JSON or CSV containing joint landmarks, the calculated angles, and a simple missing-phalanges Boolean for each finger segment.
Phase 2
Package the model behind an easy interface where I can drop in one or many photos of the same patient and immediately see the angle values plus a visual overlay confirming the landmarks. A lightweight web dashboard or cross-platform desktop app is acceptable; use whatever stack gets to a usable prototype fastest.
Acceptance criteria
• Mean absolute error of joint-angle measurement ≤ 5° on a held-out test set I will supply
• Low false-negative rate for missing phalanges/digits
• End-to-end inference time under 3 s for a 2000×2000 px image on CPU
Deliverables can be handed over in GitHub or a private repo with clear setup instructions and a short readme. I can supply the anonymised images and to iterate quickly on feedback.
• calculate the flexion/extension angle at every finger joint in each image
• flag when a finger or phalanx is completely absent so that it can be recorded automatically
I expect you will combine classical computer-vision preprocessing with a deep-learning model (PyTorch or TensorFlow are fine) and possibly leverage pose-estimation libraries such as MediaPipe or OpenCV for landmark detection. Accuracy matters more than perfection in lighting or background, so robust data-augmentation and domain-adaptation techniques will be key.
Phase 1
Train and validate a model that ingests a single hand photograph and returns a JSON or CSV containing joint landmarks, the calculated angles, and a simple missing-phalanges Boolean for each finger segment.
Phase 2
Package the model behind an easy interface where I can drop in one or many photos of the same patient and immediately see the angle values plus a visual overlay confirming the landmarks. A lightweight web dashboard or cross-platform desktop app is acceptable; use whatever stack gets to a usable prototype fastest.
Acceptance criteria
• Mean absolute error of joint-angle measurement ≤ 5° on a held-out test set I will supply
• Low false-negative rate for missing phalanges/digits
• End-to-end inference time under 3 s for a 2000×2000 px image on CPU
Deliverables can be handed over in GitHub or a private repo with clear setup instructions and a short readme. I can supply the anonymised images and to iterate quickly on feedback.