As the blog mentioned, each cell predicts three things -
- bbox coords (tx,ty,tw,th)
- objectness score (po)
- class scores (p0 - pc)
and again each cell predicts three boxes. Hence you get that big red tensor as model output.
So yes, other cells can detect "dog" and more precisely, each cell can detect any class.
what do you think the first cell (0,0) would detect?
Any object. for eg a sample one box prediction of that cell would be like -
box coords - [ 0.23, 0.12, 0.4, 0.2]
objectness score - (0.6) # this is a probability whether this box contains an object
class scores - [0.2,0.3,0.5] # lets say you have three classes, so here you will get the probability scores for those classes.
And each cell predicts three boxes, so you will get two more similar predictions from this cell. All of these are packed in that red tensor.