Extending deep learning model to detect more objects

How can we extend pretrained models (like vgg16, vgg19, yolo,etc) to detect more classes of objects. Do we have to train the model with data-set of all previous classes( plus your own class) or just with data of class which you want to add in the existing model. And if we just train on the new class dataset then should we use just convolutional pretrained model or the final pretrained model (which predicts x classes already).
In short, how can we incrementally add classes to our object detection model?


@smitsheth Please have a look at this paper. (Learning without forgetting) https://arxiv.org/abs/1606.09282

Hi @smitsheth, @akshai
have you solved this issue or have you find some examples that you can share?
Thank you