search

pjreddie / darknet

Convolutional Neural Networks
http://pjreddie.com/darknet/
Other
25.84k stars 21.33k forks source link

loss function? #854

Open yuxuan11 opened 6 years ago

yuxuan11 commented 6 years ago

loss

I want to ask you a question，what is Ci? Is the category number of the object?

kdatta03 commented 6 years ago

Ci is the confidence score. Each bounding box predicts a confidence score, along with 4 coordinates.

i is the index of the grid cell, and theoretically there should be a confidence score for each bounding box, not each grid cell. But in the expression above, only one bounding box will be considered for each grid cell during training ( the bounding box which has the highest IOU with the ground truth box). (1ij obj) will be zero for all bounding boxes in a cell except one of them (it will have a value of 1 for that box) . So since only one bounding box per grid cell - only one confidence score per grid cell. Thus, that's what Ci represents.

Let me know if that made sense :)