Figure 19. YOLO object detection algorithm training.
Despite the fact that YOLO developers suggest training YOLO over 2,000 iterations for each class the training stopped earlier. This decision was made due to the fact that the neural network reached a loss parameter which was not getting any lower even after a while. As a result, YOLO created a weights file which is used during image processing. 4.2.2 YOLO object detection To modify the touché list, we need to launch YOLO for each frame and find which objects it has detected (Figure 20). YOLO creates a text file with the resulting object detection in each frame. In some cases, it is possible to see more than one fencing bout in the frame because of the camera position in the fencing hall. So, we need to filter all the recognised objects. The solution is to identify the coordinates of each controversial object and only keep those with lower coordinates. For convenience, the filtered objects are sorted by 25