Online ring sizing looks simple from the user’s side: take a phone photo, place a known object beside the hand, and receive a ring size recommendation. The engineering problem is less simple. A useful system has to convert pixels into physical units, isolate the correct finger, measure a soft and non-circular body part, correct systematic image-measurement bias, and still behave predictably when users take imperfect photos.
The AI Ring Sizer series documents that progression from controlled computer vision prototype to a more production-oriented measurement pipeline.
- 🚀 Live Demo: sizer.femometer.com
1. Sub-Millimeter Finger Measurement
Ring Sizer, Sub-Millimeter Finger Measurement via Computer Vision
The first report covers the original local CV pipeline. It uses a standard credit card for scale calibration, MediaPipe landmarks for hand structure, anatomical zone selection for the ring-wearing position, and Sobel-based edge refinement for sub-pixel width measurement.
The core result: under controlled capture conditions, the pipeline reaches sub-millimeter measurement precision and produces a practical two-size recommendation range.
2. Calibration and Bias Correction
Ring Sizer Calibration Report, Systematic Bias Correction via Linear Regression
The second report explains why raw CV measurements are not enough. Even with stable scale detection and repeatable edge measurement, the system consistently over-estimates finger diameter compared with caliper ground truth.
This study quantifies that bias across 60 measurements, fits a linear regression correction, and validates the model with leave-one-person-out cross-validation. Calibration reduces mean absolute error from 1.58 mm to about 0.60 mm.
3. Foundation-Model Segmentation
Ring Sizer 2.0, Foundation-Model Segmentation for Real-World Photos
The third report moves the system beyond lab-style photos. The main failure cases were not measurement precision failures, but detection failures: cluttered backgrounds confused card detection, and shadows near the finger edge confused classical segmentation.
Ring Sizer 2.0 replaces the most brittle detection stages with prompt-based Segment Anything segmentation while keeping the pipeline local, CPU-friendly, and fast enough for an online demo.