TY-Shi / FreeCOS

FreeCOS: Self-Supervised Learning from Fractals and Unlabeled Images for Curvilinear Object Segmentation (ICCV2023)

FreeCOS

Pytorch implementation of the paper Accepted by ICCV 2023. "FreeCOS: Self-Supervised Learning from Fractals and Unlabeled Images for Curvilinear Object Segmentation", Accepted by ICCV 2023. In this paper, we propose a novel self-supervised curvilinear object segmentation method that learns robust and distinctive features from fractals and unlabeled images (FreeCOS). To the best of our knowledge, FreeOCS is the first self-supervised learning method for various curvilinear object segmentation applications.

Usage

1.generate the curvilinear object images

# generate the curvilinear object images in Data/XCAD/make_fakevessel.py
python Data/XCAD/make_fakevessel.py #make_fakevessel.py is an example python script.

2. Training scripts

CUDA_VISIBLE_DEVICES=0 python train_DA_contrast_liot_finalversion.py 
#(CUDA_VISIBLE_DEVICES=0 python train_DA_contrast_liot_DRIVE_finalversion.py for DRIVE)

3. Evaluation scripts

CUDA_VISIBLE_DEVICES=0 python test_DA_thresh.py

4. Trained models

Trained models can be downloaded from here. [Google Drive] [Baidu Drive (download code: 3w1a)].
Put the weights in the "logs/" directory.

5. Trained Data

Trained Data can be down from here. [Google Drive] [Baidu Drive (download code: 3w1a)] (you can generate different curvilinear data by our method or use the same generated curvilinear data as our experiment. Different generated curvilinear data will effect the performance)

Future Work

-FreeCOS will be continuously updated.

-Thanks for the parts of LIOT codes from "Local Intensity Order Transformation for Robust Curvilinear Object Segmentation" (https://github.com/TY-Shi/LIOT), we changes the LIOT to a online way in FreeCOS codes.

Contact

Thanks for your attention! If you have any suggestion or question, you can leave a message here or contact us directly:

• [e-mail]
• [e-mail]