[Paper
] accepted for CVPR'24.
Two options:
task = {query image, support image(s), binary support mask(s)}
DEMO
OR git clone
the huggingface repo
to either (a) call from_model(task)
in app.py
or (b) run the gradio app locally to let it use your GPU.Prepare the dataset: data/README.md.
Call
python main.py --benchmark {} --datapath {} --nshot {}
,
for example
python main.py --benchmark deepglobe --datapath ./datasets/deepglobe/ --nshot 1
Available benchmark
strings: deepglobe
,isic
,lung
,fss
,suim
.
Default is quick-infer mode.
To change this, pass --adapt-to every-episode
.
To turn on post-processing, pass --postprocessing [always|dynamic]
.
To change other parameters, check the available parameters in core/runner.py makeConfig()
.
Select --verbosity 1
to get printed what's currently happening while runnning the loop.
Consult eval/README.md for notes on reproducing results.
This work might give you inspiration to try some adaption before comparison for CD-FSS. You might be interested in my opinion that
If this work finds use in your research, please cite:
@article{herzog2024cdfss,
title={Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation},
author={Jonas Herzog},
journal={arXiv preprint arXiv:2402.17614},
year={2024}
}