RF-DETR with step learning rate scheduling and optimized hyperparameters
We fine-tuned RF-DETR using a step learning rate scheduler on a custom dataset. Within two epochs, the model achieved a +3.7 increase in mAP@50:95, with balanced improvement in classification and localization losses. EMA weights consistently outperformed standard parameters, indicating stable convergence. Per-class analysis shows strong performance on well-represented categories like two-wheelers and trucks, while smaller or visually ambiguous classes such as minibuses remain challenging, suggesting future improvements via data balancing.