Prototype
Experimental Race-Bib Photo Search
A local AI-assisted pipeline that associates visible race numbers with event photographs so gallery visitors can search by bib number.
How it works
When I prepare a supported event gallery, a local tool looks for people, bicycles, visible bib regions, and plausible number readings. A race-bib-specific YOLOv7 detector is combined with YOLO11 person detection, EasyOCR, and Qwen2.5-VL review. The resulting numbers are saved with each photograph and can be used for exact-number filtering in the static gallery.
The processing runs on my own computer. Photographs and recognition data are not sent to a third party for this analysis.
Limitations
This is an early-stage prototype and its accuracy has not been evaluated. It can miss small, blurred, angled, or partly hidden numbers, and it can occasionally read a number incorrectly. The search results are therefore useful hints rather than a guaranteed complete record of every participant visible in an event.
Perfect recognition cannot be promised for any event or recognition method. Different methods work best at different events, depending on bib placement, number design, lighting, and whether the numbers are worn by people or attached to bicycles. The approach may therefore need to be adapted for each event.