search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.68k stars 7.96k forks source link

Changing the network architecture and better anchor IoU algorithm #1901

Open gnoya opened 5 years ago

gnoya commented 5 years ago

Hi, I want to change the architecture (add and/or remove some layers) so I can compare the results for my particular project.

1) How would you recommend to do it? What is the thing that I need to be most aware of? 2) How would you recommend to compare results? Train in the exact same way with same cfg parameters?

Also, I made an algorithm trying to improve anchor boxes' average IoU and got about 3.7% better than the kmeans method you use in ./darknet detector calc_anchors. I'm gonna test these news anchors and see if there is an improvement.

Cheers!

AlexeyAB commented 5 years ago

@gnoya Hi,

  1. Just see the difference between yolov3.cfg and yolov3-tiny.cfg. Add more convolutional layers and may be add more yolo-layers.

  2. Use the same dataset and parameters for training different models, and compare mAP.