<![CDATA[Object detection models are pivotal for retail automation but require vast annotated datasets, which are costly and time-consuming to acquire. This creates a significant challenge in few-shot learning scenarios where data is scarce, leading to models with poor generalization. This study investigates strategies to overcome data limitations by training a YOLOv8 object detection model on a custom dataset of 152 retail product images across 28 classes. We conduct a comparative analysis of three training protocols: (1) a baseline model trained on the original data, (2) a model enhanced with advanced data augmentation techniques, and (3) a model supplemented with synthetically generated data. Performance is evaluated using mean Average Precision (mAP@50−95), Precision, and Recall. The synthetic data approach significantly outperformed the other methods, achieving the highest mAP@50−95 of 0.699 and the highest Recall of 0.856. While the data augmentation model yielded the highest Precision (0.875), its lower Recall (0.714) resulted in a suboptimal mAP. Furthermore, training with synthetic data demonstrated markedly faster and more stable convergence. Our findings indicate that for few-shot object detection in specialized domains like retail, supplementing training with synthetic data is a more effective strategy than relying solely on traditional augmentation.]]>
Copyrights © 2025
