Fresh Fruit Bunch Detection Performance: A Comparative Evaluation of Training Strategies

In this study, we evaluate the impact of different training methodologies on the performance of YOLOv5s, a popular object detection model, by comparing training time, GPU memory usage, and detection accuracy. We explore four distinct approaches: single batch training, multiple batch training, single batch training with concatenated images, and multiple batch training with concatenated images. The objective is to determine whether using concatenated images can reduce training time while assessing its effects on GPU utilization and detection accuracy, all while using the same dataset.Our experiments reveal that training with concatenated images significantly reduces the overall training time. Interestingly, GPU memory usage remains consistent with that of traditional training methods without concatenation. However, this approach leads to a slight reduction in detection accuracy compared to the native methods that do not involve image concatenation.In conclusion, for applications where reducing training time is crucial and a marginal decrease in accuracy is acceptable, training with concatenated images presents a more efficient alternative to conventional methods.