Can a smartphone photo replace calipers for glove sizing?

HandAnthro is a system that reads 44 hand dimensions from a single picture. It needs only a palm-up hand placed on a sheet of US letter paper.

How the pipeline works

The user lays a hand flat on standard US letter paper and takes a photo with a phone. The software first reconstructs the paper edges hidden by the wrist so it can rectify the perspective. It then whitens every pixel that does not belong to the hand. A fine-tuned YOLO pose model proposes 41 anthropometry-specific landmarks. Those landmarks are refined using image-specific geometry and contours before the final 44 projected dimensions are computed. The whole chain runs on the captured image without extra hardware. No specialized lighting. No trained operator. Just the phone and the paper.

Controlled test results

Forty-five held-out participants each supplied 16 captures. The captures spanned two smartphones. They also spanned two backgrounds. Two angles and two nominal illumination settings completed the matrix. That yielded 720 images. HandAnthro produced complete outputs for 704 of them. A success rate of 97.8 percent.

Against two trained operators using calipers the mean absolute error was 3.80 mm per dimension. Non-thumb fingers averaged 2.48 mm error. Thumbs averaged 6.04 mm. Palm and wrist averaged 6.17 mm. Errors vary by region.

Pilot with firefighters

A researcher-assisted mobile-app pilot processed 268 retained firefighter images screened by the research team. Automated batch processing returned all 44 dimensions for 260 images. A 97.0 percent completion rate. An unpaired descriptive comparison with an independent national firefighter reference produced a mean absolute difference of 2.40 mm across 28 sex-by-dimension group-mean contrasts. The paper describes this as feasibility evidence for future distributed studies. The comparison is not a paired accuracy study.

What the paper does not claim

The authors do not report injury data or user discomfort. They do not assert clinical validation for medical devices. How many of the 720 controlled captures remain in active use is not stated. The supplementary material spans 14 pages but the main text stops at seven.

The sources this article was written from

  • "arXiv:2609.37855v1
    https://arxiv.org/abs/2609.37855v1"

Read next