SoccerNet_v3_H250/

This is our proposed dataset for evaluating long-shot Football (Soccer) player and ball detection models. It is a subset of SoccerNet-v3. It consists of the frames in which the length (height) of the bounding box that locates a person does not exceed 250 pixels. The SoccerNet v3 compressed .png images have been extracted to the proper directory structure and converted to .jpg. Additionally, the corresponding annotations of the bounding boxes have been filtered and converted into a format compatible with YOLO annotation. Two classes are present:

• “0” for ball, and
• “1” for person bounding boxes. As "person" we have included all 7 classes of human annotations.

The division into training, validation and testing parts has been preserved, and the corresponding number of images is now: 14,368/2,726/2,692. Relevant code is available at [https://github.com/kmouts/FootAndBall]

More info regarding the dataset is available here: [https://github.com/kmouts/SoccerNet_v3_H250]