The idea of “Co-occurrence Matrix” for object classes can be extended to address some aspects of Active Learning. Suppose that we have a large portion of unlabeled data, we obtained predictions from the model and are interested to quickly identify “bad predictions”. We can click on the cells of a “Co-occurrence Matrix” and see corresponding images and predictions.
Once we identify images with bad predictions, we might want to add these images to our training set. For that, we need an easy way to tag them so that we can extract them later.
An image below illustrates the idea.
Essentially, it’s pretty much the same App as ‘App to show “Interactive coexistence matrix” for spatial objects on images’, except tagging functionality is added