search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.64k stars 7.95k forks source link

About YOLOv3-SPP #4413

Open May-forever opened 4 years ago

May-forever commented 4 years ago

Dear @AlexeyAB ，

Sorry for bothering you.

When I compare the Yolov3.cfg and Yolov3-spp.cfg, I find that you use spp structure to replace the 78 layer of Yolov3.cfg for creating Yolov3-spp.cfg. i.e., you use spp structure to replace a 3*3 conv layer in Yolov3.cfg.

I'd like to know why you use spp structure to replace the 78 layer? why you did't replace any other layer? Is there any theory support?

Looking froward to hearing from you, thanks in advance.

Best regards, May

AlexeyAB commented 4 years ago

https://github.com/AlexeyAB/darknet/issues/2859

May-forever commented 4 years ago

2859

Dear @AlexeyAB ,

Thank you very much for your help.

However, yolov3-spp.cfg uses 5x5 9x9 13x13 in SPP, while in the article 5x5 7x7 13x13: 'DC-SPP-

YOLO_ Dense Connection and Spatial Pyramid Pooling Based YOLO for Object Detection'

why you choose 9x9 instead of 7x7 ? Is there any theory support or mathematic support?

Looking forward to hearing from you, thanks a lot.

Best regards, May

AlexeyAB commented 4 years ago

This was chosen by Joseph. I don't know why, may be to increase receptieve field, may be he tested both approaches, but may be just a typo )