docs/next_steps.md

Next Steps

Work through this list in order. The project already runs locally on synthetic images; the remaining work is about proving it on real Braille and preparing a public demo.

For the full checklist, see TODO.

1. Docker Verification

Status: done locally.

docker compose up --build starts, http://localhost:7860/health returns OK, and sample upload through /upload returned hello.

Keep this command set for future verification:

docker compose up --build
curl http://localhost:7860/health
docker compose down

2. Real Braille Images

Status: partially done. Ten public real Braille photos are in ignored data/raw/, and the current pipeline has been smoke-tested against them.

Done when uploading real Braille photos through the UI produces correct or near-correct text.

• Save 20 to 30 real photos in data/raw/.
• Try close-up embossed paper, Braille book pages, labels, and signs.
• Upload them through the app one by one.
• Record output text versus expected text.

Tune src/braillevision/config.py one value at a time:

• Lower blob_min_area if real dots are missed.
• Raise blob_min_area if noise is detected as dots.
• Lower blob_min_circularity if embossed dots appear elongated.
• Raise clahe_clip_limit if dots are too faint.
• Check estimate_dot_spacing output if cells are grouped incorrectly.

3. Annotated Test Set

Done when at least 10 real images have annotation JSON.

Example file in data/annotations/:

{
  "image": "filename.jpg",
  "text": "the actual english text",
  "conditions": {
    "lighting": "normal",
    "blur": 0,
    "rotation_deg": 0
  }
}

4. Augmentation

Status: done for the current 10-image local set. data/processed/ has 80 generated variants.

Done when data/processed/ contains variants for the real-image set.

python scripts/augment_data.py --input data/raw/ --output data/processed/

The script creates eight variants per image.

5. Benchmark

Status: blocked until annotation JSON exists.

Done when README has a benchmark table with real numbers.

python scripts/benchmark.py --testset data/processed/

Record character error rate by condition and convert to accuracy:

accuracy = 1 - CER

6. Public Repo and CI

Status: done.

GitHub repo is public at https://github.com/sriksven/braillevision, Actions is green, and the CI badge is in README.md.

7. Public Demo

Status: done.

The Hugging Face Spaces deployment is live:

https://sriksven-braillevision.hf.space

Verified on May 31, 2026:

• /health returned OK
• / returned HTTP 200
• /upload with data/samples/hello.png returned hello

Browser TTS should still be checked manually in the deployed UI.

8. Demo Assets

Done when the README and submission have visible proof.

• Record a 30 to 60 second GIF for docs/demo.gif.
• Record a demo video under 3 minutes.
• Include TTS audio in the demo video.
• Show real image upload and text output.

9. Submission Notes

Be explicit and honest:

• Synthetic tests pass.
• Real-image accuracy depends on collected image conditions.
• Grade 2 support is partial.
• Upload validation is minimal.
• Report real benchmark numbers once available.