Answers are grounded in this paper and cite their sources.
Acquiring real-time, accurate, large datasets is crucial and time-consuming for specific problems. Numerous datasets are available with annotations, but most are not feasible for a special task because of differences in the class label, class imbalance, and variability. One such solution to this problem is to use artificially crafted datasets (or synthetic datasets), which are scalable and can be automatically annotated. We utilized two different approaches—stable diffusion and cut-paste-blend—to generate a synthetic dataset. This study investigates the use of synthetic image datasets to observe the performance of YOLOv8 and improved YOLOv8 models for helmet detection. We trained models on both real-world and synthetic datasets and evaluated their performance in terms of detection accuracy. After training 50 epochs, the model achieved a mAP@50 of 78.6% on real data, 45.5% on synthetic data, and 75.4% on hybrid datasets. We analyzed how the hybrid dataset affected results using different ratios and discovered that with a 3:1 mix of hybrid data, the YOLOv8-based model reached an mAP@50 of 90.3%, which is better than when real and synthetic data were used in equal amounts. We proposed the Convolutional Block Attention Module-based YOLOv8-CBAM to enhance the accuracy of helmet and non-helmet detection. Experimental results indicate that YOLOv8-CBAM achieved an mAP@50 of 91% at 50, which is 0.7% better than the baseline model. This study also indicates that the correct proportion of synthetic datasets solved the class imbalance problem and improved the helmet detection accuracy in challenging environments.
No takes yet. Share an insight, caveat, or question.
Arshad et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: