Hi there, I would like to create a video classiftication system for cctv video analysis, focused on a specific theme, my idea is to use a dataset of class1t/class2 videos to get a model that will find if there is class1 or class2 category in the videos, that is similar to the tutorial that you have “fine tuning SOTA videos model”, I mean, I will use Resnet or similar to create a new model that clasifies in two categories, good. Now it comes my doubt, I would like to use this classification model to create new models for different cctv installments, but i dont want to loose the initial categories, I want to add the new categories but without training again in the original videos that will take too much time and money, is it possible to do transfer learning without loosing classes/categories previously learned?
Adding new categories without retrain the model is a hot research topic, but unfortunately, I don’t think there is a good solution for that. For now, I think the easiest thing would be
(1) change last dense layer from 2 to 3 or more classes
(2) train the model on all of your data including previous categories. In order to save computation and money, you can freeze most of your model (e.g., the first three res blocks res1, res2 and res3 in a typical resnet). Since your data are all CCTV videos, freezing the early layers shouldn’t effect much.