I have been trying to look for help on something similar to weighted resampling but in the multi-label object detection scenario for my highly imbalanced dataset.
My dataset consists of images that are broadly classified as good images and bad images (images with defects). Furthermore, the bad images are further classified into seven different (labels 1 through 7) based on the type of defects they have. The good ones which do not have any defects are encoded as all zeros [0,0,…,0] and an image with defects 1 and 2 for example is multi-hot encoded as [1,1,0,…,0] .
The good ones are a lot in number and the ones with defects are very less. This calls for weighted re-sampling. I have looked at a discussion thread here: Unbalanced data in multiple object detection and here: Help for Sampling training data for multi-label classification task?!. Was wondering if anyone has a sample implemented. Thanks.