Clarification on Timestamps for Foul Event Clips in Dataset

[mmiakashs]

In the paper, it has been mentioned that "For both classification tasks, we leverage clips of 16 frames, spanning temporally for 1 second, with a spatial dimension of 224×398 pixels. Specifically, the clips contain 8 frames before the foul and 8 frames after the foul." However, the data annotation lacks timestamps indicating when the foul occurred. Could you please share the details of how the 1-second clip was extracted?

[heldJan]

Hello. Each clip is 5 seconds long with 25 frames per second, totaling 125 frames. The point of contact typically occurs at the 75th frame. We trimmed the clips using --start_frame 63, --end_frame 87, and --fps 17 to capture only the frames where the foul takes place, but feel free to adjust the values to your needs.