Squintly

Pairwise psychovisual data collection for zensim. Browser-based, viewing-condition-aware, coordinates with coefficient.

Existing public IQA datasets (KADID-10k, TID2013, CID22) fix viewing conditions and bake them into the labels. zensim plateaus around SROCC 0.82 on these — the residual is dominated by how an image is being viewed: device pixel ratio, intrinsic-to-device ratio, viewing distance, ambient light, gamut. Squintly collects pairwise judgments with those conditions recorded, so zensim can learn to condition on them.

Quick start (local)

# 1. Have a coefficient store available (HTTP or filesystem path)
just build
cargo run -- --coefficient-http http://localhost:8081 --port 3030

# Or:
cargo run -- --coefficient-path /path/to/coefficient/benchmark-results --port 3030

# 2. Open http://localhost:3030 in any browser, on any device. Calibrate, rate a few.

# 3. Export when done:
curl http://localhost:3030/api/export/pareto.tsv     > pareto_human.tsv
curl http://localhost:3030/api/export/thresholds.tsv > thresholds.tsv
curl http://localhost:3030/api/export/responses.tsv  > responses_raw.tsv

Deploy to Railway

railway init --name squintly
railway volume add --mount-path /data
railway variables --set "SQUINTLY_COEFFICIENT_HTTP=https://your-coefficient-host"
railway up --detach

Full walkthrough in DEPLOY.md. The Dockerfile builds the embedded TS frontend, then the Rust binary, then ships a minimal debian:bookworm-slim runtime — no Node at runtime.

What gets collected

Per session (stable):

• devicePixelRatio, screen size, color gamut, dynamic range, OS/browser
• Optional self-reported: viewing distance, ambient light, vision-corrected, age bracket
• Optional credit-card calibration → CSS px per millimeter on this physical screen

Per trial (variable):

• Viewport size (handles orientation flips on phones)
• Image intrinsic, displayed-CSS, device-pixel dimensions
• intrinsic_to_device_ratio — the headline number
• Dwell time, swap count, zoom usage

Output

pareto.tsv is in the zenanalyze/zentrain training schema, with quality set to a Bradley–Terry-fitted scalar score derived from pairwise preferences (anchored at 100 = reference). Drops in as a replacement for the zensim column.

See SPEC.md for the full design.

Status

v0.1 — works locally pointed at a coefficient store. No remote deploy.

License

Apache-2.0 OR MIT