Een master bestaat niet alleen uit dingen bouwen, maar ook uit dingen onderzoeken. Binnen de Research Methods for Artificial Intelligence module van de Open Universiteit wordt dat nog eens extra benadrukt. Binnen deze module heb ik gewerkt aan een onderzoek naar de relatie tussen de performance van YOLO modellen en de grootte van de dataset in de context van startnummerhekenning.
Hieronder vind je het abstract van het artikel.
This study investigates the relationship between training dataset size and the per formance of YOLO-based race bib detection models. Models were trained on the RBNR and TGCRBNW datasets using varying amounts of training data and evaluated us ing precision, recall, mAP50, and mAP50-95. Linear, logarithmic, and power-law regression models were fitted to characterize the relationship between dataset size and performance.
The results show a consistent positive relationship between training data size and model performance across both datasets and all evaluation metrics. However, no re gression model consistently outperformed the others, and no statistically significant differences were found after Bonferroni correction. These findings suggest that in creasing dataset size generally improves race bib detection performance, but do not provide conclusive evidence for a specific functional relationship between dataset size and performance.
Enkele bestanden die je kan downloaden.