I've been going over the YOLO paper again and I was wondering something about the loss.

Yolo divides an image into grids and then for each grid can have multiple bounding boxes to detect multiple objects whose centers fall into the same grid.

My issue with this is that there can be an issue with ordering.

For instance let's say within a grid, there are two object centers. Let's say the ground truth is the vector (a,b) where 'a' is the coordinates and class for object A and 'b' is for object B.

If the model outputs the vector (b,a) it is essentially right but due to ordering it will be punished.. is there anything that addresses this possible issue within the paper or newer versions of YOLO?

For instance DeTR uses the Hungarian method to get a bipratite matching between predictions and the closest ground truth label.

Edit: I think I found my answer within the paper, enter image description here

So before calculating the loss, yolo does do a matching between predictions and ground truth boxes. The prediction with the highest Intersection over Union (IoU) is chosen the the box "responsible" for that detection and the loss is done between that prediction and the ground truth.



You must log in to answer this question.

Browse other questions tagged .