A face shape classifier that keeps returning “oval” may be treating oval as a residual class: the label that collects faces which do not clearly show the traits of the other shapes. That is a plausible explanation for one published classifier, not a diagnosis that applies to every model. Before you change the architecture, check how the labels are defined, where the errors fall for each class, how the data is split, and what the preprocessing pipeline does to the input.
What a residual class means for face shape
Consumer face-shape taxonomies usually use oval, round, square, heart, diamond, and oblong. These are stylistic conventions, not objectively bounded natural groups. In a 2026 DEV Community article, Theo Marsh explains the pattern this way: several labels are tied to distinct traits, while oval may be described by the absence of the traits that would make a face another shape. If a face lacks a clearly angular jaw, a wide forehead, or the length pattern of another label, it can fall through to oval, even though no line of code names oval as the default.
As an Amazon Associate I earn from qualifying purchases.
That makes oval a residual category by construction. The classifier does not need a bug to produce it. It only needs a label scheme in which the other five labels require positive evidence while oval accepts whatever is left over.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What the reported classifier actually did
The classifier in Marsh’s article is a measureface tool. It measures four lengths plus the jaw angle and compares them with shape prototypes. He tested it on 43 synthetic faces. The figures below describe that one classifier on that one set; they are not population statistics.
#1 Best Overall
- Face Reading Essentials Face Shapes
- ABIS BOOK
- JY Books Sdn Bhd
| Measure | Reported value | Scope and qualification |
|---|---|---|
| Synthetic faces tested | 43 | Generated by an image model; none belong to a real person (Marsh, 2026) |
| Classified as oval | 15 of 43 | Same classifier and synthetic set |
| Classified as oblong | 4 of 43 | Same classifier and synthetic set |
| Returned two labels (paired output) | 8 of 43 | All eight pairs included oval: four oval/round, three oval/heart, one oval/diamond |
| Cases where forehead width ruled out oval | 16 of 43 | Article’s ruling-out analysis for this classifier |
| Cases where jaw ruled out oval | 16 of 43 | Article’s ruling-out analysis for this classifier |
The paired outputs matter more than the raw count of oval labels. When a classifier’s second-best answer is almost always oval, the other shapes are being decided by thin margins, and oval is absorbing the uncertainty. Marsh describes the skew as “not a flattering thing for us to publish about our own classifier.” That sentence is taken from the article’s indexed text; the full page was not independently re-checked for this piece.
What these numbers cannot tell you
- They do not measure how common oval faces are. The article states it found no peer-reviewed prevalence data for the six styling categories. Do not read the 15-of-43 result as a biological or demographic distribution.
- They are not an independent evaluation. The article reports one author’s classifier on generated faces.
- They do not prove a single cause. Data imbalance, label definitions, landmark features, and decision boundaries can all contribute. The sources reviewed do not establish that one of them explains every oval-heavy classifier.
- No regulator or standards body has addressed the issue. The sources reviewed contain no official statement on face-shape labeling accuracy.
- Corroboration is limited. A separate technical note on facial-shape classification reports variability in categorization, but only its abstract was available for this article. It supports the view that categorization reliability is an open question, and nothing more specific.
How to diagnose your own classifier
Work through these checks in order. The first two usually separate a label-design problem from a model problem.
Rank #2
1. Write operational criteria for every label
For each class, write the measurable condition a face must meet to receive that label, such as a minimum jaw angle or a forehead-to-cheek width ratio. If oval’s definition is mostly “none of the above,” the class is residual by design. You can keep that scheme, but you should document it and treat it as a modeling choice with known limits.
2. Read the confusion matrix and per-class metrics
Overall accuracy can hide a collapsed class. One public example repository reports random-forest accuracy of 0.46 and oval recall of 0.30 on a balanced 1,000-image test split. That is one repository’s result, not a general benchmark, but it shows how a modest headline number can coexist with weak per-class behavior.
A residual oval class has a specific signature to look for. Oval recall will often be high, because oval absorbs many inputs, while oval precision will be low, because many faces from other classes are labeled oval. Check the column of the confusion matrix where oval is the predicted label, not only the row of true oval faces.
3. Check for duplicates and identity leakage
If the same person or near-identical image appears in both training and test sets, scores will look better than the model deserves. A face-shape preprocessing study reports auditing near-duplicate images and identity leakage. It limits its performance claims to the dataset it studied, which is the right scope for your own results too.
Rank #4
4. Hold preprocessing constant when you compare
Cropping, alignment, rotation, and augmentation all change the geometry the classifier measures. A rotated or loosely cropped face can shift a jaw or forehead measurement enough to move it out of its intended class. The same preprocessing study describes separate preprocessing variants and argues that alignment comparisons need appropriate controls. When you test a change, keep the split and evaluation protocol identical, and vary one preprocessing step at a time.
Free tools Windows power users keep installed
One-click scans. No signup required.
5. Audit how inputs are rejected or passed through
One implementation explicitly rejects images with no face, multiple faces, or side-profile faces, and documents its alignment and cropping before classification. That is a useful pattern. If your pipeline silently passes poorly aligned or partial faces to the classifier, those inputs may land in the fallback class. Count how many inputs are rejected, how many are corrected, and how many reach the model unchanged. This does not prove the same mechanism operates in other systems, but it is a cheap check.
6. Expose uncertainty instead of a single label
If your classifier produces scores, show the top alternatives or a confidence gap rather than a single definitive label. This is a design recommendation inferred from the paired outputs described above, not a feature every face-shape tool offers. A near-tie between oval and another shape is more honest when the interface says so.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Comparing implementation choices
Several implementation approaches appear in public work, and their reported results are not directly comparable.
| Approach | What is reported | Source and limits |
|---|---|---|
| Landmark-feature classifiers | Traditional classifiers benchmarked against Inception v3 | Repository description; not peer-reviewed replication |
| Image-based CNNs versus random forests | Outcomes reported differently across experiments | Public repository; test split and metrics vary |
| Crop, alignment, and augmentation variants | Separate preprocessing variants compared, with control requirements for alignment | Preprocessing study; results limited to its dataset |
| Input rejection rules | No-face, multiple-face, and side-face images rejected before classification | Single implementation; no rejection rate reported |
To compare these options fairly, train and test each on the same split, then report per-class recall and precision, confusion patterns, variation across repeated random seeds, input rejection rates, and performance on an external dataset. Do not compare headline accuracies from different repositories as though the tests were interchangeable. A change of architecture alone is not established to fix residual-class behavior; the label definitions and preprocessing usually need to change alongside it.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
