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

How to improve mAP when I train tiny-yolo #4908

Open TomasTrnkaPLC opened 4 years ago

TomasTrnkaPLC commented 4 years ago

Hi. How I can improve a mAP during training? I run darknet with tiny-yolo config according to the tutorial. I have only 12 classes but, that is not a problem. https://imgur.com/a/YocVr1T my mAP is 71% after 6000 iterations. But that is a maximum what I reach. How I can improve that? Some special settings in cfg or?

stephanecharette commented 4 years ago

A few ideas, in the order that will probably have the most impact: 1) Review all your existing marks to make sure none of them are accidentally assigned to the wrong class https://www.ccoderun.ca/darkmark/Summary.html#DarkMarkReview 2) Make sure that all your objects are marked and you didn't skip any. https://www.ccoderun.ca/darkmark/ImageMarkup.html#MultipleObjectMarkup 3) Markup more images 4) Enable flip=1 in [net] section of .cfg (if appropriate) https://www.ccoderun.ca/darkmark/DataAugmentationMisc.html 5) Play with some of the colour settings (if appropriate) https://www.ccoderun.ca/darkmark/DataAugmentationColour.html

TomasTrnkaPLC commented 4 years ago

stephenecharette : thanks for your answer. I do not use a dark mark I just only download pascal VOC and training on that. Any other ideas about how to improve my mAP training?

PaserSRL commented 4 years ago

71% is very high! O_O