search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.79k stars 7.97k forks source link

Cfg for training both small, medium and large object #6876

Open sabrinatuli opened 4 years ago

sabrinatuli commented 4 years ago

Hi @AlexeyAB which yolov4 cfg I should use for training both small object ( approximately <10px), medium and large object ( approx >300px) together? I have seen most of the people using yolov4-tiny-3l.cfg only for small object. Also what is the difference between yolov4-tiny-3l.cfg and yolov4-tiny-custom.cfg? My goal is to get more accuracy for small object as well as keeping the large object accuracy well. Need suggestion pls.

stephanecharette commented 4 years ago

Less than 10px is unlikely to work. The recommended minimum is 16x16. Darknet and YOLO does work with slightly smaller, but I doubt you'll get a positive hit with 8x8 or 9x9.

As for .cfg, your best bet is probably as you say the three YOLO layers in *-3l.cfg. I've used it in projects that have a mix of large and extremely small. I have a video I uploaded a while ago to compare 4 different .cfg files which may be of interest: https://www.youtube.com/watch?v=p0Wn8ZNQ_uc

Also see this: https://www.ccoderun.ca/darkmark/Configuration.html

sabrinatuli commented 4 years ago

Thank you so much @stephanecharette . I will check the yolov4-tiny3l.cfg then.