Might I benefit from scale_x_y = 1.05 in tiny_yolo or tiny_yolo_pan2 ?

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

Open LukeAI opened 5 years ago

LukeAI commented 5 years ago

Might I benefit from putting scale_x_y = 1.05 in to the yolo layers of tiny_yolo or tiny_yolo_pan2 ?

AlexeyAB commented 5 years ago

Yes, add scale_x_y = 1.05 scale_x_y = 1.1 scale_x_y = 1.2 in different [yolo] layers in your cfg-file

AlexeyAB commented 5 years ago

just to know, in my cfg-files [yolo] layers go in reversed order than in the original cfg-files

LukeAI commented 5 years ago

great, I'll try it out! - If this is a "free" accuracy boost why not insert it into every .cfg ?

AlexeyAB commented 5 years ago

Because this is new feature. There are several new "free" accuracy boosts, which are not tested well yet: https://github.com/AlexeyAB/darknet/projects/1

So it will be added in new cfg-files.