Hello,

I am hoping someone can help advise me on handling the class imbalances. I am using DETR, but I am not sure how I can 'weight' the lesser classes to be higher and vice-versa for the Tin class.

Some ideas I am looking into trying -

1. train 3 binary DETR models, one for Tin class, then use transfer learning for the classes with less training examples. At inference, run each model on each image and decide on a way to choose overlapping bounding boxes.

2. train an ensemble of Tin-downsampled datasets and create upsampled augmented undersampled Other and Thatched classes.

3. just train a single Detr and weight the imbalanced classes accordingly.

Currently am stuck on 1. I modified the beginner huggingface notebook to only use Tin classes. Basically switched it to binary object detector, But I can't get it to output any bounding boxes at inference ... it just gives really low scores for them.

Any advice or help with where I should focus woudl be great. At this point, I just want to get something working.

Thanks,

Ian