Beyond Model-Centric Object Detection: Mitigating Scale Bias Through Object-Size-Aware Dataset Partitioning

Reliable object-detection evaluation is essential because reported performance depends not only on detector architecture but also on whether training, validation, and testing subsets adequately represent the object-scale characteristics of the data. However, conventional dataset partitioning often ignores object-size distribution, which can introduce scale bias and produce unstable or misleading estimates of model generalization. This study investigates whether object-size-aware dataset partitioning can improve the representativeness and reliability of detector evaluation without modifying the detector architecture. Two license-plate datasets were examined under four partitioning scenarios: natural unsorted splitting, descending size-ordered splitting, stratified size-descending splitting, and size-aware stratified sampling using strict quantized and hybrid semi-stratified variants. After partitioning, subset representativeness was assessed using Kullback-Leibler divergence and normalized Earth Mover’s Distance, while the nano variant of You Only Look Once version 11 was used as a fixed detector and trained under identical conditions at 640 × 640 resolution for 100 epochs. Performance was evaluated primarily using mean average precision over intersection-over-union thresholds from 0.50 to 0.95, supported by mean average precision at 0.50, precision, recall, and F1-score. Size-ordered splitting reduced the primary metric to 0.124 in Dataset 1 and as low as 0.005 in Dataset 2, whereas size-aware stratification recovered competitive values of 0.523 and 0.477, respectively. These findings indicate that dataset partitioning is a consequential data-level design factor and can support more dependable benchmarking for automated license-plate recognition, intelligent transportation, traffic monitoring, and surveillance applications.