Choice of Segmentation Head

[Seniab]

Hi,

What was the rational behind trying a linear, +ms, and Mask2Former segmentation head for image segmentation as the downstream task? Could a different segmentation be used instead?

Thanks!

[ayushnangia]

have you tried to code it up?
I mean the env they have given is not working. only option seems to be the Huggingface version.

[Seniab]

I have tried to code it up, but I am getting lost in how the output from the transformer backbone actually gets to the segmentation head. I was wondering if there was something special about linear, +ms, and Mask2Former, or is the pipeline from transformer backbone to segmentation head generally the same regardless of segmentation head?

[ayushnangia]

this Linear probing of DINOv2 for semantic segmentation

try this and #370

notebook might help out with the inference

for the coding part even I am stumped

[ayushnangia]

btw mind if we connect over mail?

[Seniab]

Thanks. Sure we can connect over mail.