Darknet: Google Open Images v5

[gpsmit]

Hi mr. Redmon,

First of all thanks for your awesome framework and architecture.
Im having a blast with it.
I mostly use the Google Open Images pre-trained weights that you supply on your website, its very powerful.
I believe that model was trained on openimages v4, right?
Recently Google released Open Images v5 (may 2019). Is there any chance of you retraining the yolo open images model on the updated dataset?

I would love to hear from you.

[fanzhoulll]

Hey, I am wondering do you really find the pre-trained model with openimage v4 very powerful? I tested and it seems like the performance is way worse than the pre-trained model with Coco. Have you noticed the same issue?

For example: following is the result from model trained with OpenImage

And here is the result from model trained with Coco:

[karolmajek]

YOLOv3-spp Open Images v4 is pretty good in my opinion
https://www.youtube.com/watch?v=p7oIuA3TG30

Weights from https://github.com/radekosmulski/yolo_open_images

I still need to check weights by @pjreddie

wget https://pjreddie.com/media/files/yolov3-openimages.weights
./darknet detector test cfg/openimages.data cfg/yolov3-openimages.cfg yolov3-openimages.weights

[dexception]

@fanzhoulll
Check both the datasets. Tagging is done by a 4 year old. Whole lot of issues.
If you want accuracy check each image and tag the missing parts.