CODEWITHSUNDEEP
deep-learningcomputer-visionyolo
FenryrMKIII
FenryrMKIII

Reputation: 1198

yoloV3 recomputing anchors

I know that yoloV3 uses k-means algorithm to compute the anchor boxes dimensions. As far as I have understood, the default yoloV3 anchors, namely : 

anchors = 10,13,  16,30,  33,23,  30,61,  62,45,  59,119,  116,90,  156,198,  373,326

have been computed on COCO dataset. However, I can't find back those values when performing the calculation on this dataset. I have tried 2014 and 2017 versions. I have tried train set separately and train+valid set. I have tried personal k-means algorithm and the one provided by AlexeyAB version of darknet. None of those trials succeeded in outputing the default anchor boxes provided in darknet.

I would really liked to validate the calculation and get back the orignial anchor boxes to make sure the anchors are calculated correctly. Anyone has performed this work ?

Upvotes: 0

Views: 739

Answers (1)

Nopileos
Nopileos

Reputation: 2117

I think I did the same some time ago and came to the conclusion they were tuned by hand a bit. You can have a look at the commit history in the darknet github and track the changes to the anchors in different configs. I believe there was a version where my calculations and the ones in the repo matched.

Also how much do the anchors differ. Is there huge difference between your anchors and the ones in the repo? If it is only minor things I wouldn't worry as much.

Upvotes: 1

Related Questions