search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.77k stars 7.96k forks source link

avg IOU doen't increase anymore.. #2029

Open richardminh opened 5 years ago

richardminh commented 5 years ago

Hello, AlexeyAB, How are you? I am training my own car logo data with darknet... But i can't see increasing aveIOU anymore (no more changed around 0.3XXXX ~ 0.4XXX) Why? I prepared logo dataset with cropped from vehicle image by using labelImage app..

AlexeyAB commented 5 years ago

@richardminh Hi, Im fine, thanks!

Check your dataset by using https://github.com/AlexeyAB/Yolo_mark And check does mAP increase during training? What cfg file do you use?

richardminh commented 5 years ago

Hi, AlexeyAB, thanks for your kind reply. easy.cfg file was used and i labeled image with YOLO_mark app. and i found obj:none, no_obj:none in training phase. What happened during in train?

AlexeyAB commented 5 years ago

and i found obj:none, no_obj:none in training phase.

This is normal.

Only if avg loss is nan - then training goes wrong.

Just download new version of Darknet from this repository and try to run training with -map flag and you will see mAP values at some points: ./darknet detector train data/obj.data yolo-obj.cfg darknet53.conv.74 -map

mAP (mean average precision) is the most important accuracy indicator, it is more important than IoU or avg loss.

chart_map_tiny_v3