Darknet: Why are there two calls for delta_yolo_box() in forward_yolo_layer()?

[eyalzubia]

First, thanks very much for this repository!

1. The first call to delta_yolo_box in forward_yolo_layer function is the straightforward one (near line 230). But I can't understand why there's a second call, updating l.delta, near line 261, iterating over mask_n. Could you explain?

2. Also, I don't understand where is the confidence loss - I see only class delta and box delta updates.

Thanks!

[AlexeyAB]

1. 
   
   
   
   • 1st is for any predictions which have IoU > truth_thresh, i.e. delta_yolo_box can be calculated for several final cells/anchors (but in current models truth_thresh=1 so this part of code is disabled)
     
     https://github.com/AlexeyAB/darknet/blob/6682f0b98984e2b92049e985b21ed81b76666565/src/yolo_layer.c#L222-L231
   
   

• 2nd is for best prediction only, so delta_yolo_box will be calculated for the best final cell/anchor only: https://github.com/AlexeyAB/darknet/blob/6682f0b98984e2b92049e985b21ed81b76666565/src/yolo_layer.c#L259-L261

---

1. confidence (probability) = objectness * class_probability

• objectness:

  • https://github.com/AlexeyAB/darknet/blob/6682f0b98984e2b92049e985b21ed81b76666565/src/yolo_layer.c#L216-L223
  • https://github.com/AlexeyAB/darknet/blob/6682f0b98984e2b92049e985b21ed81b76666565/src/yolo_layer.c#L263-L265

• class_probability:

  • disabled cuz truth_thresh=1 in cfg-file: https://github.com/AlexeyAB/darknet/blob/6682f0b98984e2b92049e985b21ed81b76666565/src/yolo_layer.c#L228
  • https://github.com/AlexeyAB/darknet/blob/6682f0b98984e2b92049e985b21ed81b76666565/src/yolo_layer.c#L270

[eyalzubia]

Thanks!