Revision history [back]

I would suggest retraining for a single class. It makes common sense that you do not want nodes in your network to remember classes you will never use and thus make your detector weaker for the actual class you are looking for. Retraining using the YoloV2 architecture is quite straightforward and PJReddie explains it quite well on his own website. I would suggest trying YoloV2 first, because it is actually the most powerfull architecture of the 3 versions. YoloV3 still has some issues, that you can avoid if the actual problem it tries to solve (very small objects in ratio to the image dimensions) is not occuring in your dataset!