3 ms·
YOLO is amazing. When I first looked at the loss function in YOLO v1, I was stunned: You can do regression loss(Mean square error) instead of classification los
by itissid 2y ago
YOLO is amazing. When I first looked at the loss function in YOLO v1, I was stunned: You can do regression loss(Mean square error) instead of classification loss(categorical cross entropy) on class prediction probabilities, and it can still just "work".
- wongarsu 2y agoI 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.