[06pratyush]

The eval page that docs/contracts.md, section 3 asks of the app, so eval/run_detect.py (#10) can run the real detector instead of the stub.

What is in it

• app/eval.html and app/src/eval.ts set window.solidMask:
  • decode(url): fetch, then the app's own decode() (createImageBitmap with imageOrientation: "from-image").
  • detectFaces(image, options): the app's detect.ts.
  • render(image, masks): draws the given masks (already padded) and encodes exactly as the app saves. The drawing moves out of main.ts into paint() in mask.ts, which the page and render() both call. render() rounds each mask outward to whole pixels and clips it with maskFor(mask, 0, …), as section 2 (v0.2) says. With the app's own masks, its file is byte-identical to the app's saved masked-1.jpg in all three engines.
  • info(): { ortVersion: "1.30.0", backend: "wasm", threads: 1, modelSha256 }. modelSha256 is the hash of the model as loaded: detect.ts now records it when it reads the file.
• window.solidMask appears once the model has loaded and compiled, so the runner's wait covers it and its first image is timed like the rest.
• A page that Vite reloads because a source file changed refuses to start. Without this, a run could go on under code that changed in the middle of it, all recorded under one git commit.
• npm run dev serves the page at http://localhost:5173/eval.html, the address eval/README.md gives, and now fails rather than move if 5173 is taken. The page has the app's Content Security Policy.
• The deployed build leaves the page out: Vite builds only index.html, and a test checks that no eval code is in app/dist.

Tests

tests/app/eval-page.spec.ts drives the page on the dev server, as the runner does (the image served on the page's own origin, the runner's options), in Chromium, Firefox and WebKit.

• Golden detection. The 7 boxes scoring 0.10 or more must match OpenCV 4.13's FaceDetectorYN on the same model:

  • within 0.5 px and 0.0005 of score in Chromium;
  • within 2 px in Firefox and WebKit, since WebKit on macOS decodes JPEG with ImageIO.

  The reference comes from tests/reference/opencv_yunet.py, with OpenCV pinned, and its output is committed, so you can rerun it. Swapping the colour channels makes the test fail: the face is still found, 93 px lower, so the privacy suite wouldn't notice. Nothing guarded the equivalence with OpenCV before; the review of the eval PRs flagged this.

• A photo stored turned, with the EXIF Orientation 8 the runner adds for Rotation 90, decodes upright with the face within 2 px.

• render() and info(). render() covers the face and writes a file with nothing but the picture, and info() returns exactly the expected values.

• A reloaded page refuses to start.

Playwright now starts the dev server too (port 4402), next to the static server for the production build.

Checked with the eval code

I ran eval/run_detect.py --page http://127.0.0.1:5173/eval.html and eval/score.py on a two-image stand-in manifest made from the public-domain test photo, one copy as stored and one stored turned 90°. No Open Images image was used.

• Run line: the model's SHA-256, ORT 1.30.0, wasm and 1 thread.
• Faces: the same face on both copies (within 0.3 px), each fully covered and found, with no false masks.
• Timing: median 333 ms per 12 MP image, max 395 ms. The first image no longer carries the model's compile time.

For Shashwat

• Detect reloads in the runner. run_detect.py could also stop when the page navigates mid-run (page.on('framenavigated')). The page now refuses to start after a reload, so the run fails on its next image anyway, but the runner would then say why.
• Tiling and YOLO5Face. The runner's --tiles and --model yolo5face still stop at detectFaces, which supports only the single setting and YuNet so far. Tiling comes next, for part B's tiled setting.

/code-review and two independent reviewers looked at this before I opened it, and all their findings are fixed.

🤖 Generated with Claude Code