This problem is related to the situation when a window owns several bounding boxes' centers.

As this screenshot shows here:

The bounding box represents the car, and the center of the bounding box represents a person. Looking at the structure we've seen so far, we can't solve this problem, because each of these windows has the following structure:

This structure can represent only one bounding box, so the windows in the screenshot can have either the car bounding box dimensions or the person ones, but not both.

And if we look at the grids for each of the windows, we have these structures:

We're going to solve this problem by using different shapes of bounding boxes, which are called anchor boxes.

In this case, we choose two boxes: one the fits to a car shape and one that's a bit taller than it is wide, which fits to a person. For each of these anchor boxes, we use this structure:

P_c is the bounding box properties and the class:

Now, notice how the dimensions change:

These will be the number of the anchor boxes. We've doubled the amount of information needed.

In the car and person picture, for example, the P_c for the anchor box surrounding the car would be 1, because we have an object with a rectangular shape. And the class will be 1 0 0 since this is a car.

Then we try for the shape of the bounding box surrounding the person: P_c will be 1, the bounding boxes properties will represent the list shape, and now we'll have 0 1 0, which represents the person.

In this way, by choosing different shapes, we're able to assign more information to a window, and a window basically can own many centers of the bounding boxes.