United States
Body Measurement project
Software supplier
Product
The task was to create a solution to precisely calculate anthropometric measurements from a photo person and input basic parameters. The acceptable standard error of the measurements is 5%.
The solution included using Computer Vision and Machine Learning. A person inputs their weight, height, and gender as the basic parameters and takes two full-size photos from the front and the side.
Technology Stack
- Python 3.7
- PyTorch
- Mediapipe
- Docker
- Flask
- PostgreSQL
- Angular
The algorithm works as follows:
- A body frame is detected by using CV;
- Inside the frame, a body contour is calculated using a gradient filter;
- Using Google’s Mediapipe, the body contour is mapped to a pose model;
- Cross-width of the contour is calculated at the predefined landmarks, which are mapped to the contour using the pose;
- ML regression is used for converting linear measurements into circumferences;
- ML outlier detection and regression are used to validate the calculated measurements and their improvement.
The algorithm calculates ten measurements of different parts of the person’s body: Neck, Elbow, Wrist, Chest, Waist, Hip, Thigh, Knee, Calf, and Ankle. The measurements are given as circumferences. No helper markers are used in the photos.
