nehabalamurugan/spot-the-ball

Spot the Ball Dataset

Overview

This dataset is part of the Spot the Ball research project, which explores how humans infer the location of a soccer ball in images where it has been masked. The dataset consists of thousands of labeled soccer images, processed using CLIP-based selection, YOLO object detection, and Stable Diffusion inpainting.

Dataset Structure

Pre-processed (Raw Image Data)

Folder Name	Number of Images
clipped_images	2908
clipped_images_2	749
val_clipped_images	465
validation_clipped_images	78

Post-processed (Formatted for Training)

• Grided/ – Images with a 10×6 grid overlay (700×420 resolution).
• Resized/ – Images resized to 640×360.
• Truth/ – CSV files containing ground truth labels.

Data Format

Each image has a corresponding row in the ground truth CSV file:

image_name,x_min,y_min,x_max,y_max,player_count,avg_player_area
image_0.png,323.0,247.5,351.0,274.5,4,16117.25
image_1.png,305.0,278.0,339.0,310.0,4,31229.75
image_10.png,185.5,305.5,196.5,314.5,5,1950.0