Anchor boxes predefined template boxes of certain height and width ratio. These are used in YOLO  to help detect multiple objects from a single grid cell. We define the shape of boxes based on rough geometry of type of objects that can be detected.

Currently, as explained, our model will be able to detect only one object per grid cell, but in most cases, you may have multiple objects per cell. Remember that we consider the cell closest to the object to be the center:

To solve this, we need Anchor points; basically, we will add to our output depth volume a predefined bounding box; then, during training, we choose the objects with the center closest to a particular cell and choose the bounding box with the biggest IoU to the anchor box. In practice, the idea of anchor boxes make the network generalize the detection better, due to the fact that several subnetworks will be responsible for looking for other objects on the same cell.