suppose we are presented with a folder of images. the task is just detect if any new image should belong in this folder or not.
in this folder, there may be natural groups of images that are similar to each other, but the groups themselves are very different from each other and is probably ineffective to be generalized under one class. for example maybe we have images of sports cars, then images of dogs. The images in the folder are not divided in any way. there are no labels for any of these images. except that whatever is in the folder is what we want to look for from new images. how do we approach this problem in the most efficient and effective way?
one approach i thought of was perhaps train a deep clustering model, and also separately train a few shot open set recognition model. divide the folder into N classes, then take K samples from each of the classes in the folder and check if a new image should belong to any of these. having multiple steps like this using separately trained models may accumulate too much error in the end. any other suggestions?
