Open AlasterMeehan opened 4 years ago
in this repo https://github.com/AlexeyAB/darknet yolov3.cfg training has the same bahavior currently and 2 years ago
in this repo https://github.com/pjreddie/darknet yolov3.cfg training has the same bahavior currently and 2 years ago - only if you use paramter [net] letter_box=1
and run detection with flag -letter_box
Does a model or .cfg file behave the same way from v3 to v4? What changes are there to detection and training behaviors.
From what I read in the paper it seems to me that a previously trained in v3 will give the same output in v4 but with some optimization?
Training may however behave differently due to the "bag of freebies"?
In some cases we may not use some of the features in the "bag of freebies", for example orientation may be an important part of object classification and rotation distortions would spoil this. Is there any documentation for controlling these parameters and what the defaults are?