This project is aimed at developing a low-cost solution to measuring infant height and weight. The focus is on measuring height, since this is the more challenging metric. Currently, this is a tedious process that requires an 'infantometer'. The existing analogue approach is (1) slow, (2) uncomfortable for the baby, (3) innaccurate and (4) requires multiple people.
The approach consists of two core steps. Firstly, we perform pose detection to detect keypoints of the infant. Secondly, we use a known real-world object to determine the relative size of image pixels.
By performing a vector sum along the pose vectors and scaling by our known object size we can recover the true size of the baby.
Off-the-shelf pose networks show useful keypoints. However, we could likely improve performance and accuracy by storing manually placed keypoints and fine-tuning an existing network on image poses.
The real-world object chosen is a Bicycle Poker playing card. These are a standard size, widely available and cheap. Further, they can easily be placed on the baby in such a way that (at a distance) the card seems to lie on the same plane as the baby.
baby_pics- contains input imagesopencv- initial C++ prototyping code for measuring heighttflite- Python code for pose detection and segmentationproto- mobile app (written in Flutter) for taking images, computing height and viewing growth charts
- Improve documentation
- Add tests
- Update to newer tensorflow plugin, use newer Pose model
- In-app walkthrough/tutorial
- 2nd degree polynomial line of best fit
- Add pics and overview of functionality to Wiki