SF-SAM-Adapter

SF-SAM-Adapter, a advanced segmentation method based on the observation of the spatial and frequency features of the segmented target, integrates the prior knowledge to the large segmentation model SAM, by simple yet effective Adapter technique Adaption.

Requirement

conda env create -f environment.yml

Download SAM checkpoint, and put it at ./checkpoint/sam/

Example

1. The dataset is placed in "./data/eye", and the file structure of dataset is:

EYE/

 EYE_Test_Data/...
 
 EYE_Training_Data/...
 
 EYE_Test_GroundTruth.csv
 
 EYE_Training_GroundTruth.csv

2. Train: python train.py -net sam -mod sam_adpt -exp_name *eye* -sam_ckpt ./checkpoint/sam/sam_vit_b_01ec64.pth -image_size 1024 -b 32 -dataset eye --data_path *../data*

3. Evaluation: The code can automatically evaluate on the test set during traing, and you can also manually evaluate it by running val.py for.

Results will be saved at ./logs/ in default.