Abstract:
Human Pose Estimation (HPE) is a cornerstone in the progress of computer vision, with the YOLOv8 algorithm emerging as a leading framework due to its remarkable performance. This study concentrates on improving YOLOv8’s accuracy and generalizability by integrating synthetic data into its training process. Utilizing Nvidia Omniverse, known for producing highly realistic synthetic environments, we crafted a dataset tailored for HPE enhancement. The integration of this synthetic dataset aimed to sharpen the precision of YOLOv8, broadening the scope of its applicability. Results showed a significant improvement in model performance, with an up to 19% increase in mean Average Precision (mAP) at 0.5 IOU and up to 12% rise in mAP across the 0.5-0.95 IOU range compared to models trained on conventional datasets. These findings highlight the potential of synthetic data to augment real-world data collection and underscore its value in developing more robust and precise HPE models, encouraging a shift towards innovative training approaches in computer vision.