YOLOv8 dfl_loss metric

I was wondering how to interpret different losses in the YOLOv8 model. I've easily found explanations about the box_loss and the cls_loss. About the dfl_loss I don't find any information on the Internet. I've also checked the YOLOv8 Docs.

I've found an article about the Dual Focal loss but not sure it corresponds to the YOLOv8 dfl_loss : Dual Focal Loss to address class imbalance in semantic segmentation

Could someone explain to me what is the dfl_loss and how to analyse it ?
Thanks !

There is an explanation on Matlab page: https://www.mathworks.com/matlabcentral/fileexchange/104395-dual-focal-loss-dfl?s_tid=FX_rc2_behav

Broadly speaking, DFL loss 'considers' the problem of class imbalance while training a NN. Class imbalance occurs when there is one class which occurs too frequently and another which occurs less. For ex: In street imagery say 100 photos, one can have 200 cars and only 10 bicycles. One wants to detect both cars and bikes. This is case of class imbalance, when you train a NN, since there are lot of cars, NN will learn to accurately localize cars whereas, bikes are too less so, it might not learn to localize it properly. With dfl loss, every time the NN tries to classify bike there is increased loss. So, now NN puts more importance on less frequent classes. This explanation is on a very general level. To know more, refer the paper on Focal loss and then on DFL.

DFL stands for Distribution Focal Loss. Class imbalance is not relevant. It is used for bounding box regression along with CIOU. Although I haven't fully grasped the paper yet, it seems that the calculations also take into account the vicinity of the ground truth because the ground truth for the boxes is not always completely trustworthy. To understand this, I believe studying anchor-free detection is also necessary.

https://github.com/implus/GFocal

Let's break down Distribution Focal Loss (DFL) with a simple example.

Imagine you have a model that is trying to classify images into three categories: cat, dog, and bird. Let's say you have a dataset with 100 images, but the distribution of the classes is very imbalanced. Specifically, you have 80 images of cats, 15 images of dogs, and only 5 images of birds. So, most of the images are cats, and very few are birds.

When training your model, the standard focal loss can help to give more importance to the rare classes (dogs and birds) during training, making the model pay more attention to them. However, the standard focal loss doesn't take into account how well the model's predicted probabilities match the actual distribution of the classes in the dataset.

Here's where Distribution Focal Loss (DFL) comes in. DFL not only considers the importance of rare classes but also pays attention to how well the model's predictions align with the actual distribution of the classes. In our example, DFL would encourage the model to predict probabilities that match the actual distribution of cats, dogs, and birds in the dataset (80%, 15%, and 5%, respectively).

To achieve this, DFL adjusts the loss based on the differences between the predicted probabilities and the target probabilities. If the model predicts a high probability for cats (e.g., 90%) but the actual distribution in the dataset is only 80%, DFL will give it a penalty for the misalignment. Similarly, if the model predicts a very low probability for birds (e.g., 1%) when the actual distribution is 5%, DFL will penalize this as well.

By considering both the importance of rare classes and the alignment with the target distribution, DFL helps the model to make more balanced predictions and improve its performance, especially on datasets with severe class imbalances.

Keep in mind that the actual formula for DFL might involve more complex calculations, but this simplified explanation should give you a basic understanding of its purpose. In real-world applications, the model's predictions are typically refined iteratively during training to find the best alignment with the target distribution and achieve better object detection performance.