Loss function

[Baghdad-Hamdoud]

@viplix3 Thank you for your work,

• I would like to know which loss function do you use ?

• Is it possible to add another function and where ?

[viplix3]

TL;DR
What loss function is currently being used?
For regression: Mean Squared error
For classification: Sigmoid cross-entropy

Is it possible to use other loss functions?
Obviously. Some suggestions...
For regression: Huber Loss (smooth L1 loss), likes of GIoU, DIoU etc.
For classification: I don't think so changing this can help much, sigmoid cross entropy works fine.

@baghdadhamdoud how the YOLO algorithm works is, it tries training a model which can do both, regression and classification simultaneously.
What is even the sense of doing two completely unrelated tasks simultaneously you ask?
Well, to detect where a particular object is in an image, we first need to tell if there even is an object in an image. This is where the classification part comes in.

THE CLASSIFICATION
YOLO gives a certain score to each box. Teaching a model to differentiate between an object and background can be treated as a classification task, where the two classes will be
Class 1 - Yes, some object is present.
Class 2 - No, no object is present.

Once you've identified that there is some object present, I think you would probably want to know what that object is. You don't want to pet a cat thinking it is a dog, no, NO! And this my friend is where another classification loss comes in and it tries to teach the model if there is an object present, what that object is.

THE REGRESSION
If you're still reading the crap I have been blurting out about the obvious things, I believe you really want to know how regression works, and I'll tell ya.
See, YOLO is an object detection algorithm, if we don't tell where the said object is present, it won't really be object detection. So, to get the coordinates of bounding boxes, YOLO uses regression. It tries to teach the model to predict the offsets of the mid-point of the predicted box w.r.t. some grids and along with it, the scaling factors of an anchor (an assumption made by us humans of how the box should be like), factors by which some said anchor should be stretched or squished to get a tight bounding box in the prediction.

[Baghdad-Hamdoud]

I thank you very much @viplix3 , i understand more how object detection works in YOLO.

Now i would to know, is it possible to use Focal Loss Function ?
It would be interesting to use it in view of what it can bring, it can solve the problem of class imbalance.

If it is possible, how in your repo ?

[viplix3]

Not the actual focal loss, but a special case of focal loss has been implemented and is available in the repo, as pointed out in issue #10
The catch is, not a lot of testing has been done. You can follow this link for the exact code.