In the IOU lecture we say we take the intersection but why would we have more than 1 bounding boxes of 1 object. When we train we justify a single bounding box then even if we test the dataset we might get a slightly wrong bounding box but no correct bounding box to take its intersection with.
I can think of 3 places IOU is used in YOLO. First, it is used to assign a ‘best’ anchor box shape to a labelled object while setting up the training data. Second, it is used to compare ground truth with predictions when computing error due to inaccurate coordinate predictions in the yolo loss function. Finally, it is used in NonMaxSuppression to compare two predicted boxes to determine whether they are likely to be duplicates or represent unique objects.