darknet
darknet copied to clipboard
Is it possible to have the same confidence as single class model and multi class model?
Hello, I am wondering if it is possible to achieve the same or similar confidence in detection when you train a model with just one class versus multiple classes. For example, if I train one model with just one class and train the other one with the same annotations but with other classes annotation on top of it, how does that affect the accuracy of that one particular class? Having more classes will result in more iterations so I know that's one factor of it being different, but if both models are properly trained, shouldn't they have roughly the similar results?
Having more classes makes your problem more complex, not necessarily less accurate. So the confidence can probably be as high as a single class model as long as the training parameters are set correctly. Sorry that I can't give you a definite answer.
I have done tests to train a weight file of three classes, one class appears in the image, and the recognition confidence is above 0.9. If two classes appear in the image. The recognition confidence for each class is reduced to 0.4. Even some classes are not recognized I do not know why?