I was starting out with deep learning and come across a lot of pretrained models in frameworks and sites such as tensorflow model zoo. Are these models actually used by other developers in real use cases. I could see it was possible to retrain models using transfer learning techniques and building on top of the pretrained model. But my question is more specifically would there be be applications using the pre trained models out of the box without any further retraining.
For eg a I could see models like resnet and mobilenet as common occurances in many tutorials where people do an inference with these but I am confused how to take it further. These models are trained on datasets with 1000s of classes so I am wondering how would it make sense to deploy a model trained on 1000+ classes to any custom day to day usecase. Could anybody help me with an example of a real existing application that makes use of a common model like resnet 50 out of the box without retraining?
In such a scenario how would the model be utilised would it predict one of the 1000 classes and then would there be some other algorithm to do the rest?