Research demo. The model was trained only on small (128 × 128 px) ice-hockey images. It works best on hockey photos where the number is clearly visible; on other sports, large photos or crowded scenes it will miss numbers and make mistakes. It finds where numbers are; it does not read the digits.
1. Choose an image
Drop an image here, or
Loading…
The default, 0.35, gave the best balance of hits and false alarms on the validation images.
2. Numbers found
Results appear here after you upload an image.
| # | Score | Top-left (px) | Size (px) |
|---|
The model
- JerseyNet: a CenterNet-style detector with a residual backbone and a feature pyramid down to stride 4.
- Trained from random initialisation on 3,165 images of the public jersey-2 set, with mosaic augmentation, a GIoU box loss and a 384 px input.
- Test AP50:95 0.314 and AP50 0.593 (this checkpoint); 0.308 ± 0.005 over three seeds.
How it runs
- The model is exported to ONNX and runs with ONNX Runtime Web (WebAssembly).
- The image is letterboxed to 384 × 384 exactly as in training; the boxes are decoded from the centre heatmap without NMS.
- On the test split this app's pipeline reproduces the Python results (AP50:95 0.314).
Paper and code
- Preprint: Localising jersey numbers without pretraining (Zenodo).
- Code and results: github.com/MHKhan50/jersey-number-detection (v2.1.0).