Computer vision & automation
GameRL OCR pipeline
77 of 78 items read exactly, vs 21 of 78 before
- Computer vision
- AI
Evidence
- Data
- Labelled game-screen frames, one game profile · 78 items on 13 frames
- Stage
- Prototype
- Stack
- Python, RapidOCR, Qwen2.5-VL-3B
Evaluation
What was measured, against what, and how.
| Result | 77 of 78 items |
|---|---|
| 95% interval | 93% to 100% (Wilson score interval on 77 of 78) |
| Baseline | 21 of 78 for the original stack (EasyOCR + legacy tooltip parser), same set |
| Sample size | 78 items: 14 numeric zones on 1 frame, 8 on 4 frames, 56 tooltip fields on 8 frames; one game profile |
| Timeframe | Bench runs 2026-05-01 |
| Method | bench_ocr.py: exact match per item against ground_truth.json; RapidOCR for numbers, Qwen2.5-VL-3B for tooltips |
| Source | bench/results/bench_final_check_v5.json and bench_easyocr_baseline.json in the GameRL folder (local) |
| Confidence | Verified |
Limits
- One game profile and 13 frames
- Exact match only: no character error rate; latency logged but never summarised
- The one miss is a small '4' glyph the detector skips
- The reinforcement-learning part of GameRL was planned and never started
Outcome: Succeeded. Goal: read the game's numbers and tooltips reliably. 77 of 78 labelled items were read exactly, against 21 of 78 for the original stack. The reinforcement-learning part never started.
What was built
The deliverables this project covers.
- GameRL OCR pipeline tool Prototype
- Scanned-card OCR comparison report Prototype
How I used AI
What the tools did, and what was checked.
Built with: Jett in claude.ai and Claude Desktop chats (incl. claude.ai Design).
Note to be written: tool and model, what the assistant produced, what was rejected or rewritten, and which check caught a problem.