Skip to content

LEARNINGS.md

Latest commit

 

History

History
70 lines (49 loc) · 2.11 KB

LEARNINGS.md

File metadata and controls

70 lines (49 loc) · 2.11 KB
Raw

Bib Tagger - Technical Learnings

Architecture Decision: ONNX-Only Stack

We chose ONNX Runtime over PyTorch/PaddlePaddle for distribution size:

Stack Size
PyTorch + PaddlePaddle ~2GB+
ONNX Runtime ~250MB

This makes the app practical for distribution as a standalone macOS .app or .dmg.

OCR Model Selection

Model Versions Tested

Package OCR Model Year Accuracy
rapidocr-onnxruntime 1.2.3 PP-OCRv3 2022 Baseline
rapidocr 3.4.3 PP-OCRv4 2023 +5% bibs read
PaddleOCR (native) PP-OCRv5 2024 +10% but 4x slower

We use rapidocr 3.4.3 which bundles PP-OCRv4 ONNX models internally. This provides a good balance of accuracy and speed without external model files.

RapidOCR 3.4.3 API

from rapidocr import RapidOCR
ocr = RapidOCR()
result = ocr(image)
# result.boxes - list of polygon coordinates
# result.txts - tuple of recognized text strings
# result.scores - tuple of confidence scores

Height Filtering for Secondary Numbers

Race bibs often have multiple numbers:

  • Main bib number (large)
  • Gear check numbers (small)
  • Timing chip codes (small)
  • Wave/corral indicators (small)

The 50% height threshold in read_bib_number() keeps only text regions at least half as tall as the largest detected text. This reliably filters secondary numbers without needing to train a separate classifier.

Performance Benchmarks

On a test set of 782 race photos:

Metric Value
Average time per image ~420ms
Detection rate 92% of images with visible bibs
OCR accuracy ~95% on clear bibs

Future Improvements

PP-OCRv5 Conversion

PaddlePaddle models can be converted to ONNX for better accuracy:

pip install paddle2onnx
paddle2onnx --model_dir PP-OCRv5_det --save_file models/PP-OCRv5_det.onnx

This would provide PP-OCRv5 accuracy with ONNX speed, but requires maintaining custom model files.

Training Data

The bib detection model was trained on public Roboflow datasets. See README.md for full attribution.