OpenCV 4 with Python Blueprints - Second Edition

Training SSD- and YOLO-like networks

You are already aware that networks such as YOLO and SSD predict objects with predefined anchor boxes. Out of all available boxes, only one box is chosen, which corresponds to the object. During prediction time, the box is assigned with the class of the object and the offsets are predicted.

So, the question is, how do we choose that single box? You might already have guessed that IoU is used for that purpose. The correspondence between the ground truth boxes and anchor boxes can be made as follows:

1. Create a matrix that contains all IoU values of all possible ground truth and anchor box pairs. Say, the row corresponds to the ground truth box and the column corresponds to anchor box.
2. Find the maximal ...