Marathon Photo Finder

Machine learning system that helps runners find their race photos automatically using person re-identification and bib-number OCR.

2024-03 View on GitHub
Python PyTorch CNNs OCR Flask OpenCV

Marathon Photo Finder solves a common frustration for runners: sifting through thousands of official race photos to find the handful where they appear.

Motivation

Race photographers capture tens of thousands of images per event. Runners currently scroll through entire galleries sorted only by kilometre marker. This project makes the search instant.

How it works

  1. Person detection — A CNN-based architecture locates every runner in each photo.
  2. Bib OCR — An optical character recognition system reads the race number from detected bounding boxes, providing a precise identifier for re-identification.
  3. Re-identification — Given a runner’s bib number (or a reference photo), the system returns all matching images from the event gallery.

Key challenges

  • Bib numbers are partially occluded, crumpled, or motion-blurred in race conditions.
  • Varying lighting and crowd density require robust detection at different scales.