I imagine the YOLO researchers would argue MSE is more "correct" in their case. Yolo has augmentations that combine images with different labels. If you make an

I imagine the YOLO researchers would argue MSE is more "correct" in their case. Yolo has augmentations that combine images with different labels. If you make an image that's to 30% composed of the image of a train and to 70% the image of a car, you can argue it should be predicted as 0.3 train, 0.7 car, 0 everything else. But now you are suddenly doing a regression task.