Clarify SegFormer-B0 evaluation caveats
#2
by masteroshi006 - opened
Document that the IGN test evaluation uses positive images only and that the published result comes from a single seed. The checkpoint, run config, and metrics are already present from merged PR #1.
Thanks @masteroshi006 ,the documentation fixes are merged, and thanks again for the SegFormer-B0 checkpoint.
One thing to keep in mind for users: the model takes 416×416 inputs (BDAPPV’s 400×400 images are resized to 416 because SegFormer needs dimensions divisible by 32), so images should be resized to 416 before inference.
Next step is on my side: I’ll document how to access the models and add SegFormer-B0 to the benchmark.
gabrielkasmi changed pull request status to merged