Building a segmentation model for table tennis racket using synthetic data

Table tennis players attempt to control the ball's motion by adjusting the movement of the racket. Multiple studies have been conducted on the analysis of racket manipulation by measuring the motion of table tennis rackets. However, these studies have employed methods that require attaching markers to the racket, and none have analyzed racket manipulation during actual matches. In this study, we explore a method for extracting segments containing table tennis rackets from images using deep neural networks. Segmenting the racket is useful as a preprocessing step for racket pose estimation. YOLOv8, a state-of-the-art object detection and segmentation model based on the YOLO architecture, was adopted for detection and segmentation. We created a synthetic dataset using Blender. Additionally, we extracted relevant images from COCO dataset that contain objects commonly visible in table tennis match videos. Finally, we combined these datasets for training YOLOv8. To enhance the robustness of racket detection and segmentation, we applied data augmentation, including image blurring, color transformation, random channel reordering, and brightness and contrast adjustments, ensuring that the model learns to recognize rackets under various lighting and visual conditions. The trained model was evaluated on a validation dataset consisting of 1000 images. The results demonstrated that the model achieved a precision of 0.952, recall of 0.856, mean average precision at IoU 0.5 (mAP50) of 0.915 for segmentation. These results indicate that the proposed method effectively detects and segments table tennis rackets in images, providing a foundation for further applications in pose estimation and match analysis.