Face shape detector accuracy: why results can differ

Understand camera distortion, pose, hair, landmark limitations and why a profile-similarity score is not an accuracy percentage.

There is no validated accuracy percentage here

We have not evaluated this shape classifier against a representative, independently labeled dataset. It would be misleading to put a success rate or a confidence percentage beside a result. A successful landmark detection does not prove that the styling label is correct.

The result page does show relative profile-similarity bars so you can see nearby shapes. Photo results use ranked labels instead of percentage confidence. A longer bar means a closer fit to our proportion rules. It does not establish how often that label is right.

What the three photo checks establish

The analyzer measures the original image, a padded version, and a mirrored version. It combines the middle value of each ratio and checks whether the readings agree. If the outline changes too much, or no profile fits, it asks for a clearer photo. When nearby shapes overlap, both names appear prominently.

Agreement across these views tests repeatability within one photo. It cannot establish the true label, recover a hidden hairline, or correct every camera distortion. The profile centers were initially adjusted using seven generated illustrations. Testing those same illustrations is a software regression check, not independent evidence of face-shape accuracy.

What published research tells us

Tio’s Inception v3 study considered five classes using 500 celebrity photos. Its author explicitly cautioned that generalization needed a larger dataset and proper cross-validation. Those reported scores do not validate this website’s different seven-class method.

A 2025 study of Pöch’s facial-shape system found substantial overlap and poor agreement between observers in its sample. That study concerns a different classification system and population, but it shows why a category should not be presented as a universally exact measurement.

A photograph changes the outline

Camera distance, lens perspective, head tilt, rotation, expression, and lighting all affect visible proportions. A smile can change the cheek outline. Facial hair can hide jaw corners. Hair, hats, heavy shadows, or frames may cause the landmark model to infer points that are not clearly visible.

Our checks reject missing or multiple detected faces, small or clipped outlines, severe darkness or washed-out exposure, strong blur, obvious head turns or tilts, and large mouth openings. These checks have engineering thresholds; some unsuitable photos may still pass. A relaxed expression and level camera remain essential.

Use an estimate responsibly

1. Take a clear, level photo with the whole face visible and no beauty filter.
2. Check whether the overlay lines match the visible outline. If an endpoint looks wrong, discard the estimate.
3. Compare the description with your own observation. You may find that two profiles fit.
4. Choose styling ideas based on your preference and practical needs, not the category alone.

Model limitations and representation

We do not have evidence that performance is equal across all ages, skin tones, facial characteristics, camera types, or accessibility needs. A failed detection means the tool could not interpret that image; it says nothing about the person in it. The manual calculator and quiz offer alternatives, but their categories are also subjective.

Before publishing an accuracy rate, we need real photos with independently agreed labels, all seven categories, and people kept separate between development and testing. Evaluation must count retake requests as well as completed results, report results for each shape, and disclose uncertainty and sample size. That validation has not yet been completed.

When to skip the label

If repeated results feel confusing or unhelpful, use the style finder directly. You can compare frames, hair silhouettes, and makeup placement without measuring anything. These tools are for everyday styling curiosity, not medical decisions, identity verification, or evaluating someone’s worth.

Learn how the analyzer makes a suggestion or explore styles without a photo.