I don't get it why, when having a 3x3 grid, it can still detect an object that overlaps 2 grids?

[ai_curious]

@Alireza_Saei

The Yolo network predicts adjustments to predefined anchor boxes of varying dimensions

+1 for calling out this really important point that isn’t always mentioned or emphasized. In YOLO v2, the predicted bounding box shape is the anchor box shape times a factor. The derived bounding box shape is used in the cost function, but the factor is what the network is learning to generate. This also reinforces why it is important to have a good set of anchor boxes; the closer the anchor box shapes are to the ground truth bounding box shapes, the faster and better the network learns the weights that produce better factors and lower localization error.

@Zolids there are many threads that go deep on the questions you pose above. Here is one that might be useful:

Applying YOLO anchor boxes

Applying YOLO Anchor Boxes A previous thread provides an introduction to extracting from labelled ground truth data a small set, on the order of 3 <= k <= 10 , of shapes for what is referred to in the YOLO 9000 paper as dimension clusters or priors, but in this class as anchor boxes . NOTE: Just to make things more confusing, the YOLO 9000 paper uses anchor boxes to refer instead to the hand-picked shapes described in the Faster R-CNN paper, not the ones generated by K-Means. The YOLO 9000 pap…

It contains the mathematical expressions for what @Alireza_Saei wrote and that I quoted above. From those equations you can see how a predicted bounding box shape can exceed the dimension of a grid cell.