Darknet: yolov3.conv.105 vs yolov3.conv.81vs darknet53.conv.74

[takuhara]

I am trying to learn the 2 class problem.
One is person and the other is a minor class which is neither in imagenet nor coco.

In yolov3, I know the following three pre-trained models,
Which do you think is the best in this case?
The reason for this is also?

yolov3.conv.105
yolov3.conv.81
darknet53.conv.74

[AlexeyAB]

Try to use yolov3.conv.81

[baolinhu]

where can i get the yolov3.conv.81 or yolov3.conv.105？I only find darknet53.conv.74. thanks.

[takuhara]

no download link.see below.

https://github.com/AlexeyAB/darknet/commit/ba70801e982ada241ced55bedad2a411da1896c4#diff-2a86661590614a8bbb5321cf28e9ed97

[AlexeyAB]

@baolinhu

Download file: https://pjreddie.com/media/files/yolov3.weights

How to get file yolov3.conv.81 https://github.com/AlexeyAB/darknet/blob/eff487ba3626a39e135d13929117e04bc4cf5823/build/darknet/x64/partial.cmd#L21

[lvshuaigg]

What is the difference between yolov3.conv.105 and yolov3.conv.81.Thanks.

[AlexeyAB]

@lvshuaigg

• darknet53.conv.74 takes 74 layers from darknet53.weights
• yolov3.conv.105 takes 105 layers from yolov3.weigths
• while yolov3.conv.81 takes 81 layers [0 - 80] - this is correct pre-trained weights.

yolov3.conv.105 is incorrect - because layer-81 depends on number of classes, so higher layers will be shifted if number of classes will be changed.

[lvshuaigg]

Thank you very much.

1. If I only train one class, what's the difference between yolov3.conv.81 and darknet53.conv.74?
2. What is the difference if I train a lot of classes?

[AlexeyAB]

@lvshuaigg
You can use any of them: yolov3.conv.81 (is Trained on MS COCO) or darknet53.conv.74 (is trained on Imagenet)

• darknet53.conv.74 takes 74 layers from darknet53.weights
• yolov3.conv.81 takes 81 layers from yolov3.weigths

[kmsravindra]

@AlexeyAB , Can I use darknet53.conv.74 for 832 x 480 aspect ratio? If darknet53.conv.74 is obtained from imagenet (I understand the image aspect ratio here is mostly 1:1) / Pascal VOC data, does it reduce the accuracy when I train the custom images of 832 x 480 aspect?

[AlexeyAB]

@kmsravindra Yes, you can. darknet53 is trained with crop-data-augmentation (like jitter in the Yolo): https://github.com/AlexeyAB/darknet/blob/21a4ec9390b61c0baa7ef72e72e59fa143daba4c/cfg/darknet53.cfg#L13-L14
So it is slightly invariant to aspect ratio.

I think it will be better to use darknet53.conv.74 than not to use darknet53.conv.74 in your case.

[kmsravindra]

Thanks for your reply @AlexeyAB.

1. I assume that the min_crop and max_crop are cropping the original dimensions of the image (maybe both width wise and height wise)?
2. I also meant to ask if you are aware of any other pre-trained weights for yolov3 that work better than darknet53.conv.74 for 832 x 480 aspect ratio.