A grid cell is one rectangular region of an image prediction grid. YOLO assigns object-center responsibility to cells and predicts box coordinates, size, confidence, and class information at those locations. section 12.2.3 section 12.2.4
A cell participates in detection through several values:
- Its x and y coordinates express the object center relative to the cell, from 0 to 1. section 12.2.3
- Its width and height express the predicted box size relative to the whole image. section 12.2.3
- Its confidence is 1 for an actual labeled box and 0 for an empty target location. section 12.2.3
- Its classification and box errors are weighted differently because most cells contain no object. section 12.2.4
Converting a box to a cell
For a normalized box, compute its center, multiply the center coordinates by the grid size, and take the integer coordinates as the cell index. Store the fractional center offset together with the width and height. Converting back combines the cell index and offset, then reconstructs the image-relative box.
Visualizing every cell makes the representation concrete: a class activation map marks labeled regions, while each cell's confidence and decoded box can be drawn over the image.
