Jumabek Alikhanov

https://github.com/Jumabek linked from  https://groups.google.com/forum/#!topic/darknet/qrcGefJ6d5g

anchors
https://github.com/Jumabek/darknet_scripts/blob/master/gen_anchors.py#L52

Willie Maddox
Thought I'd share my version of the anchor points calculation:

https://gist.github.com/WillieMaddox/3b1159baecb809b5fcb3a6154bc3cb0b

The main difference here is the use of fast numpy array ops. With k = 5 and roughly 40k bboxes, It only takes 1 second to calculate anchors for pascal. The coco train and val datasets take about 16 seconds. I used some parts of Jumabek and Luiz code so thanks to them for getting me started.

links
Yolo