Localising jersey numbers without pretraining: how much of the gap to YOLOv8n is pretraining?

We study how far a detector trained from random initialisation can go in localising jersey numbers in ice-hockey images, and how much of its gap to a COCO-pretrained YOLOv8n is due to pretraining. Our earlier single-box regressor barely beat a constant mean box (test mean IoU 0.044 against 0.016). We replace it with JerseyNet, a 4.7M-parameter CenterNet-style detector trained from scratch. We compare it with YOLOv8n trained from scratch and from COCO weights on the same public split. Every model is scored with COCO-style AP over all boxes, with bootstrap intervals and three seeds. JerseyNet reaches AP50:95 of 0.241 ± 0.006. YOLOv8n reaches 0.352 from scratch and 0.384 from COCO weights at the same 256 px input, and 0.419 at 640 px. Pretraining is worth about 0.03 AP, the input size another 0.035, and the detector design and training recipe 0.111. Four further experiments locate that gap: Trained with JerseyNet's light augmentation, YOLOv8n from scratch falls to 0.280 ± 0.012, so augmentation accounts for about two-thirds of the gap. Mosaic, a GIoU loss and a 384 px input raise JerseyNet to 0.308 ± 0.005. This halves the gap to a from-scratch YOLOv8n at the same input size (0.366 ± 0.005). At that input size, both detectors have the same AP on the smallest numbers. With 10% and 25% of the training images, COCO pretraining still adds only 0.04 AP. JerseyNet loses half its accuracy at 10%, against a quarter for YOLOv8n. On this task the gap to a modern detector is mainly a recipe gap, and above all an augmentation gap, rather than a pretraining gap. Code, configurations, Kaggle notebooks and archived results: https://github.com/MHKhan50/jersey-number-detection (version 2.1.0).