Pre-trained weights for custom object detection using yolov3

Question

Deep-learning newbie here, I am trying to build a custom object detection model using yolov3. After reading some documentation/tutorials, I found that is often suggested to use pre-trained weights used for other datasets such as ImageNet or COCO, even when your custom data/labels have no relationship with these datasets. Is this true (if so why)?

I would like to do object detection on a very specific type of images, specifically screenshots of websites. Should I use pre-trained weights even in this case or is it better to do the training completely from scratch?

Answer 1

This sounds like something you would probably have to experiment with, but my past experiences and community knowledge says its worth a try. You can check this stack post for more clarification, but it essentially gives the network a head start, allowing it to sort of "fine tune" itself, thus elimination some annoying qualities that would otherwise make themselves known if you were training the model from scratch.

https://stats.stackexchange.com/questions/193082/what-is-pre-training-a-neural-network