Darknet: bounding box coordinates, random detection, wrongly configured cfg for training

[maharleeeka]

1. I wanted to auto crop the objects detected, how do I get the coordinates of the boxes from predictions.png? - https://github.com/AlexeyAB/darknet/issues/912#issuecomment-392251258

2. Is it normal for weights in 300,400,500... to not detect anything at all, even if it's random detection?
   the current avg loss is 2.3 something -> https://github.com/AlexeyAB/darknet/issues/912#issuecomment-392262118
   configuration:

3. yolov2-voc.cfg
4. filter = 45
5. classes = 4
6. batch = 64
7. subdivision = 32

[AlexeyAB]

May be it helps you: https://github.com/AlexeyAB/darknet/issues/846#issuecomment-390785676

[maharleeeka]

@AlexeyAB Thank you. Do I have to build darknet in VS again after I edit some files inside src folder?

[AlexeyAB]

Yes.

[maharleeeka]

@AlexeyAB out of topic question (as I don't want to make a new issue), is it normal for weights 300,400,500... to not detect anything at all even if it's just a random detection, no accuracy at all?

[AlexeyAB]

Yes. Until 2000 iteration it shouldn't detect anything.

[maharleeeka]

@AlexeyAB at iteration 900, it actually detected the objects I wanted (though not all) but not that accurate, is this normal?

also, I noticed I actually used yolov2-voc.cfg copy for my training (yolov2) instead of what you mentioned in your docs which is yolo-voc.2.0.cfg, will this work fine or should I just cancel and retrain using yolo-voc.2.0.cfg? Thank you very much for answering my questions :)

[AlexeyAB]

@maharleeeka It's okay, training is going well.