Skip to content
🔬

Methodology

We transparently document the technical principles and data-building process behind Hogamdo's face analysis results

Last updated:

🧰 Core Technology Stack

Hogamdo relies on the following proven open-source and official AI models:

  • Google MediaPipe Face Landmarker
    An open-source face landmark detection framework released by Google. It extracts 478 3D facial points in real time. 공식 문서 →
  • ArcFace (DeepFace 라이브러리)
    A deep learning model that converts a face into a 512-dimensional vector (embedding) to measure similarity. It is one of the industry-standard face recognition models in wide use. DeepFace GitHub →
  • In-house Scoring Algorithm
    Combines 11 facial metrics (face ratio, eye size, cheekbones, lips, jawline, etc.) with per-country embedding data to compute affinity scores for 139 countries.

⚙️ Analysis Pipeline

1. Face detection and landmark extraction

MediaPipe detects the face in the uploaded photo and extracts 478 3D landmark coordinates (x, y, z). This step runs locally in your browser, so the original image is never sent to our servers.

2. Facial metric calculation

From the landmark coordinates we compute 11 facial-ratio metrics: face aspect ratio, eye-size ratio, inter-eye distance, nose ratio, lip thickness, jawline angle, cheekbone width, facial symmetry, and more.

3. Embedding-based similarity

The 512-dimensional embedding vector extracted by ArcFace is compared against each country's pre-built average embedding using cosine similarity.

4. Per-country scoring

Metric similarity and embedding similarity are combined as a weighted average to compute an affinity score (0-100) for each of the 139 countries. Countries within the same region receive additional calibration so regional characteristics are reflected.

📐 Facial Metrics Glossary

Definitions of the six core metrics used in the analysis results and statistics pages. All values are ratios computed from landmark coordinates and represent relative proportions within the face, not absolute sizes.

📏 Face Ratio

The ratio of face height to width. Higher values mean a longer face, lower values a rounder one.

👁️ Eye Size

The area of the eyes relative to the whole face. Higher values mean the eyes occupy a larger share of the face.

🔺 Jaw Sharpness

How sharply defined and angular the jawline contour is. Higher values mean a more defined outline.

🪞 Face Symmetry

How symmetric the left and right facial landmarks are, measured from 0 to 1. Closer to 1 means more symmetric.

👃 Nose Ratio

Nose size relative to the face, derived from nose length and width combined.

👄 Lip Fullness

Lip thickness relative to the face. Higher values mean fuller lips.

🌐 Data Coverage

Hogamdo's per-country average-metric dataset is currently built at the following scale. All figures are aggregated averages and do not identify any individual.

139
Countries Analyzed
100,521
Face Images Analyzed
4,815
Sample Groups Collected
View full per-country data statistics →

📊 Data Sources

The per-country facial-feature database was built from the following public datasets and sources:

  • Facial anthropometric data from public demographic and anthropological literature
  • Average embeddings aggregated from public face-image datasets for each country (individuals not identifiable)
  • Public research and survey data on aesthetic preferences in each cultural region

The data is aggregate information representing each country's 'average characteristics' rather than any specific individual's face, and is improved and updated regularly.

⚠️ Limitations and Caveats

To help you interpret Hogamdo's analysis results correctly, we state the following limitations:

  • 'Affinity' is a reference indicator based on each country's 'average tendency' and does not represent any individual's actual preference.
  • Standards of beauty are subjective and change with era and culture. This service is provided as entertainment reference material, not a scientific verdict.
  • Shooting conditions such as lighting, angle, expression, and resolution can affect the analysis results.
  • It must not be used as grounds to interpret any ethnicity, country, or appearance as superior or inferior.

📜 Algorithm Changelog

v8g63 (2026-04)

지역별 편향 튜닝 완료 — 8개 지역 전체에서 이상치 편차 15.2pp까지 축소

v8 (2026-03)

ArcFace 임베딩 도입으로 기하학 메트릭 단독 방식 대비 정확도 향상

v1 (2026-02)

MediaPipe Face Landmarker 기반 초기 모델 출시 (139개국)