Volkus Scan — on-device demographic estimation demo

Live scan

Camera permission is requested only when you press this.

Idle — models are not downloaded until you start.

◯

camera off

—

Estimated age

—rolling avg

bracket —

Gender estimate

—

— confidence

Skin tone

—

Ethnicity estimate

—

— confidence

Scan telemetry

readings 0 · stability — · elapsed 0.0s

Privacy

No frames are uploaded. Video stays in the page as a MediaStream; pixels go straight from the <video> element into a canvas and then into WASM memory.
No data is stored. This demo writes nothing to localStorage, no cookies, no IndexedDB. Reload and every reading is gone.
Models run locally via onnxruntime-web compiled to WebAssembly — inference happens on your CPU, in this tab.
Zero network calls after model load. The four .onnx files are served same-origin from ./models/; once the browser has them, the scan makes no further requests.
No fingerprinting. This page is camera-only — no device fingerprint, no KV store, no server endpoint. It is fully static.

Third-party bytes: the MediaPipe FaceMesh runtime and the ONNX Runtime WASM binaries are fetched once from cdn.jsdelivr.net. Those are code downloads — no image or result data is ever sent anywhere.

How it works

1

MediaPipe FaceMesh tracks 468 landmarks per frame and gives a face bounding box (padded 12%).

2

The box is cropped to a 224×224 RGB patch. A slightly wider crop (+22%) is used for ethnicity, which benefits from hair and jaw context.

3

Pixels become float32 CHW tensors, ImageNet-normalised (mean .485 .456 .406, std .229 .224 .225).

4

Four ONNX models run per reading: race_fp32 (7-class), gender_fp32 (2-class), age_regr_fp32 (scalar regression), and age_fused_fp32 — a regressor conditioned on a 9-dim one-hot of the predicted race + gender.

5

Skin tone is sampled separately in plain JS: a 3×3 pixel average at four forehead landmarks (10, 67, 297, 9), classified into a Fitzpatrick-style zone via an ITA-weighted score.

6

Readings are collected at ~2/sec and rolling-averaged over the ~10-second scan window. Stability is 1 - σ/15 on the age series — a single frame is noisy, the average is not.

Model payload is roughly 35 MB total (10.2 MB fused-age, 8.9 MB age, 8.9 MB gender, 8.9 MB race). They are static same-origin files, so the browser caches them — the first scan pays the download cost, later scans start immediately.

This is one layer of something bigger

Volkus is live at https://volkus.net — an anonymous social experiment where this demographic layer travels with every post.

Open volkus.net ↗ ← Back to projects

Volkus Scanon-device demographic estimation

Live scan

Privacy

How it works

This is one layer of something bigger

Volkus Scan
on-device demographic estimation