Mask2Former for Binary segmentation

[manuCeron96]

Hey. I’m doing binary segmentation using Mask2FormerForUniversalSegmentation.from_pretrained and Mask2FormerImageProcessor. I’m following this tutorial Transformers-Tutorials/MaskFormer/Fine-tuning/Fine_tuning_MaskFormerForInstanceSegmentation_on_semantic_sidewalk.ipynb at master · NielsRogge/Transformers-Tutorials · GitHub

My custom dataset contains a single class. I’m not quite sure if Mask2Former supports binary segmentation or should I treat this task as a 2-class segmentation, considering 0-background and 1-myClass.
I have tried both cases but I’m struggling with the arguments ignore_index and reduce_labels.

The model does not learn and the IoU is around 0.2 for both classes (background and myClass) along all epochs.

I’m stuck in this point, any idea?

Thanks.

[beschmitt]

Maybe a long shot but have you ever found out the correct settings for binary segmentation? I’m also trying to train mask2former for a binary task with 0-background and 1-myClass and I just don’t understand the different settings in the model and the image processor.
My model is always just returning a null-tensor no matter what I do and the documentation on the model and image-processor are practically non-existent.