Datasetpublic

headshapeai/infant-head-proportion-measurement-agreement

Photo-Based Measurement Agreement for Infant Head Proportion Aggregate agreement statistics between an automatic photo measurement of infant head proportion and human-reviewed reference values, across 3,349 photographs spanning January 2024 to July 2026. Three rows, 28 columns. No photographs, no per-photograph records, no identifiers — this is aggregate data only. Why this exists We could not find an equivalent figure published by anyone offering this kind of… See the full description on the dataset page: https://huggingface.co/datasets/headshapeai/infant-head-proportion-measurement-agreement.

sourceHugging Facecc-by-4.0updated 13d agoView on Hugging Face

0likes42downloads

Dataset Card

Photo-Based Measurement Agreement for Infant Head Proportion

Aggregate agreement statistics between an automatic photo measurement of infant head proportion and human-reviewed reference values, across 3,349 photographs spanning January 2024 to July 2026.

Three rows, 28 columns. No photographs, no per-photograph records, no identifiers — this is aggregate data only.

Why this exists

We could not find an equivalent figure published by anyone offering this kind of measurement. Accuracy is asserted often in this space and quantified rarely, and a measurement whose error is unknown is difficult to use well. So we published ours, including the part that does not flatter us: 12.0% of photographs are measured more than five points off.

Headline figures

From the primary window (n = 3,349):

Metric	Value
Three-way agreement (long / typical / wide)	85.9%
Mean absolute error, width-to-length	2.46 points
Median absolute error	0.56 points
Recall on clearly visible left–right differences	80.5%
Photographs off by more than 5 points	12.0%

The last row is the one worth carrying. Roughly one photograph in eight is measured poorly, most often where the traced outline picked up bedding, hair or clothing along with the head. An automatic outline-quality check flags 27.5% of photographs and catches 69.9% of those large deviations — useful, but not a substitute for a plain background.

What is measured

Two proportions are read from the traced outline of a head in one photograph taken from directly above, and each is compared against a reference value for the same photograph, reviewed by an experienced team.

Quantity	Definition
Width-to-length ratio	Outline width ÷ outline length, ×100. Conventionally the cephalic index; a ratio of 0.80 is an index of 80.
Left–right difference	Difference between two diagonals of the outline taken 30° either side of the vertical axis, ×100.

Dataset structure

One row per analysis window. All error figures are on the ×100 scale of the quantity concerned (cephalic-index points), not percentages of the value.

Column	Meaning
`window`	Which analysis window this row summarises
`unique_photos`	Photographs contributing to this row
`ratio_r`	Pearson correlation, automatic vs reference, width-to-length
`ratio_mean_error` / `ratio_median_error`	Mean / median absolute error, width-to-length
`ratio_bias`	Mean signed error (automatic − reference); positive = automatic reads higher
`ratio_within_{1,2,3,5}_pct`	Share of photographs within 1 / 2 / 3 / 5 points
`lr_r`	Pearson correlation, left–right difference
`lr_mean_error` / `lr_median_error`	Mean / median absolute error, left–right difference
`lr_within_{1,2}_pct`	Share within 1 / 2 points, left–right difference
`three_class_agreement_pct`	Agreement on a three-way read of the ratio (thresholds 75 and 85)
`large_lr_reference_n`	Photographs where the reference recorded a clearly visible difference (≥ 3.5)
`large_lr_both_n` / `large_lr_agreement_pct`	How many of those the automatic measurement also recorded, and the share
`large_lr_ours_only_n`	Recorded ≥ 3.5 automatically where the reference did not
`deviation_over_5_n` / `deviation_over_5_pct`	Photographs off by more than 5 points
`mean_error_excluding_deviations`	Mean absolute error over the remaining photographs
`detector_flag_rate_pct`	Share flagged by the automatic outline-quality check
`detector_recall_pct` / `detector_precision_pct`	That check's recall and precision on large deviations
`large_deviation_clean_outline_pct`	Large deviations whose outline nonetheless fits an ellipse well — failures the check cannot see
`note`	Short description of the window

The three windows

`primary_2024-05-07_onward` (n = 3,349) — the headline. One photograph per unique image, a single reference convention throughout, every reference value human-reviewed.

`excluded_before_2024-05-07` (n = 910) — published rather than discarded. Excluded from the headline because the reference measurement's own method was still being adjusted in that period: the reference median shifts by 3.0 points as an abrupt step around 2024-05-06/07, while the automatic measurement's own median is identical in both windows (81.9) and so is its segmentation quality (ellipse-residual median 0.0247 in both). The evidence points at the reference side rather than the measurement side — but the exclusion is a judgement call, so its numbers are here for anyone who wants to weigh it.

`all_windows` (n = 4,259) — both combined, published so the effect of the exclusion can be seen rather than taken on trust.

Uses

Suitable for: comparing photo-based anthropometric measurement against a human-reviewed reference; as a reference point for error rates in consumer measurement tools; as an example of a published first-party evaluation including failure rates.

Out-of-scope use

—Not a clinical validation. It compares two measurements of a photograph. It says nothing about health, development, or what any value means for an individual child.
—Not a population reference. This is a convenience sample, not balanced by age and not randomly drawn. The distribution describes the photographs measured, not the prevalence of any proportion among children.
—Reference values were reviewed by an experienced team, not by independent blinded assessors.
—Repeat-photograph consistency — the same head photographed twice, on different devices or in different light — is a separate question and is not addressed here.

Citation

This dataset is archived with a permanent DOI. Cite the DOI, not this page — it resolves for good, and it is the same two files, byte for byte.

doi:10.5281/zenodo.22706467

<https://doi.org/10.5281/zenodo.22706467>

Method: <https://www.headshapeai.com/methodology/> Full write-up, including two earlier revisions of this analysis and why they were wrong: <https://www.headshapeai.com/validation/> Benchmark page: <https://www.headshapeai.com/measurement-benchmark/>

Licence

CC BY 4.0 — reuse and redistribution permitted with attribution.