DCFM

The official repo of the paper Democracy Does Matter: Comprehensive Feature Mining for Co-Salient Object Detection.

Environment Requirement

create enviroment and intall as following: pip install -r requirements.txt

Data Format

trainset: CoCo-SEG

testset: CoCA, CoSOD3k, Cosal2015

Put the CoCo-SEG, CoCA, CoSOD3k and Cosal2015 datasets to DCFM/data as the following structure:

  DCFM
     ├── other codes
     ├── ...
     │ 
     └── data

           ├── CoCo-SEG (CoCo-SEG's image files)
           ├── CoCA (CoCA's image files)
           ├── CoSOD3k (CoSOD3k's image files)
           └── Cosal2015 (Cosal2015's image files)

Trained model

trained model can be downloaded from papermodel.

Run test.py for inference.

The evaluation tool please follow: https://github.com/zzhanghub/eval-co-sod

Usage

Download pretrainde backbone model VGG.

Run train.py for training.

Prediction results

The co-saliency maps of DCFM can be found at preds.

Reproduction

reproductions by myself on 2080Ti can be found at reproduction1 and reproduction2.

reprodution by myself on TITAN X can be found at reproduction3.

Others

The code is based on GCoNet. I've added a validation part to help select the model for closer results. This validation part is based on GCoNet_plus. You can try different evaluation metrics to select the model.