HeightEstimation

Code for the JMIR AI research letter Child Stunting Screening: Estimating Height From Smartphone Photographs. It estimates a child's standing height from a front and a side photograph and classifies stunting with the WHO Child Growth Standards.

Method

The camera lens and the upper edge of a paper marker on the wall are both 70 cm above the floor, so the marker row is the horizon in the photograph. Height then follows from

H = camH × (botPx − topPx) / (botPx − midPx)

where topPx, botPx, and midPx are the rows of the top of the head, the feet, and the horizon, and camH is 70 cm (64 cm when the child stands on the 6 cm weighing scale). Grounding DINO and SAM 2.1 segment the child, marker, and scale; MediaPipe Pose locates the head. The front-view and side-view estimates are averaged with equal weight.

Usage

Requires Python 3.13. A GPU is used when available.

pip install -r requirements.txt
python Height.py

Place the inputs next to Height.py:

• Data/<uuid>_front.jpg and Data/<uuid>_side.jpg: one photograph pair per child
• Data.csv: columns uuid, gender (Male or Female), age_days, height_cm_self_reported (stadiometer height), timestamp, and lat
• who_tables/: WHO length/height-for-age LMS tables (included)
• pose_landmarker_heavy.task: MediaPipe pose model (included; Apache 2.0, Google)

The script writes results.csv (id, age_days, gender, height_cm, predicted_height_cm) and prints the agreement and classification statistics reported in the paper, with bootstrap 95% CIs.

Data

The study photographs and measurements are confidential and are not included.