Darknet: Why FN is negetive number and recall>1?

[hemp110]

I use "detector map" to evaluate the pretrained yolov3.weights on a person dataset where only persons are annotated. The result is like this

Other classes are all zeros, which is reasonable, but in the end

FN is negetive number and recall>1, I guess some detections hit the same ground truth, and FN is calculated by unique_truth_count-TP.
I think it's better to record whether each GT is detected and define FN as the number of GTs that are not detected.

[AlexeyAB]

FN is negetive number and recall>1, I guess some detections hit the same ground truth

There were multiple hit for GT, so TP/FP/FN/IoU could be calculated wrong. (with mAP everything was fine.)
I fixed it. Now I check was this GT already detected or not. In any cases should be unique_truth_count=TP+FN

What TP/FP/FN/IoU can you get now?

[AlaaMaz]

I was running into the same problem, it appears to have been fixed now

[hemp110]

Yes, it goes well now, thanks!
I'm very glad to see the recall of benchmark weights decreases by nearly a half, and my new-trained weights appears to work a little :)