As the localizer, the YOLO loss function is broken into three parts: the one responsible for finding the bounding-box coordinates, the bounding-box score prediction, and the class-score prediction. All of them are Mean-Squared error losses and are modulated by some scalar meta-parameter or IoU score between the prediction and ground truth:
The member 1ij obj member is used to modulate the loss based on the presence of an object on a particular cell i, j:
- If an object is present in grid cell i and the jth bounding box having the highest IoU: 1
- Otherwise: 0
Also, 1ij noobj is just the opposite.
