Alright, let me run through a thought experiment here: Could I extend ImageNet database to include custom classes?
Suppose I wanted to add just one more class. I could add images and labels relating to my custom class and train a YOLOv5 from ground up. This will for sure take a tonne of time because now there are 1001 classes.
Is there anything that could be done here?
One solution that comes to mind, which I don’t is right or not…
Maybe fine-tune a pre-trained model on a custom dataset to train a model that can recognize the custom classes? Have just a few images from each class? Then a lot of examples for my extended class?
I’m thinking this can be more efficient than training a model from scratch on the full ImageNet dataset, and will allow one to tailor the model to your specific needs.