Darknet: Which characteristic for which structure - cfg files

[YKritet]

Hi Alexey,

I decided to keep the ratio aspect for my text images using the -letter_box flag and I wanted to ask about the differences in performance (not the mAP necessarily, more like the specific characteristics) between 4 cfg files :

   - yolov3_5l.cfg 
   - yolov3-spp.cfg
   - yolov3.cfg
   - darknet53.cfg

Thank you very much

[AlexeyAB]

@YKritet Hi,

• yolov3_5l.cfg - optimized for small and big objects
• yolov3-spp.cfg - increased +3 mAP mainly for medium and large objects

• yolov3.cfg - default Yolov3 model

• darknet53.cfg - is a Classifier backbone which is used for all of these Detectors: yolov3_5l.cfg , yolov3-spp.cfg, yolov3.cfg

[H-YunHui]

@AlexeyAB
as you say above that, the yolov3_5l.cfg - optimized for small and big objects, the yolov3-spp.cfg - mainly for medium and large objects.
I think whether the big objects, medium objects and small objects all improved that join SPP in yolov3_5l.cfg ? have you try it?