CosalPure

Official implementation of "COSALPURE: Learning Concept from Group Images for Robust Co-Saliency" in CVPR 2024.

Requirements

We recommend creating a new conda environment for this project. Conda can be installed through below instructions.

# clone our repo
git clone https://github.com/V1len/CosalPure
cd CosalPure

# create conda environment
conda create --name new_env --file environment.txt

Datasets

Used in our paper:

• Cosal2015 (50 groups, 2015 images) "Detection of Co-salient Objects by Looking Deep and Wide, IJCV(2016)''

• iCoseg (38 groups, 643 images) ''iCoseg: Interactive Co-segmentation with Intelligent Scribble Guidance, CVPR(2010)''

• CoSOD3k (160 groups, 3316 images) ''Taking a Deeper Look at the Co-salient Object Detection, CVPR(2020)''

• CoCA (80 groups, 1295 images) ''Gradient-Induced Co-Saliency Detection, ECCV(2020)''

Add degradation

For adversarial attack, please refer to the augment variant of Jadena. Check attack.ipynb for details.

For common corruption, please refer to the degradation process of ImageNet-C

Method

# concept learning
python concept_learning.py

# concept-guided purification
python concept_guided_purification.py

Evaluate

Used in our paper:

GICD"Gradient-induced co-saliency detection.ECCV(2020)"

GCAGC"Adaptive graph convolutional network with attention graph clustering for co-saliency detection.CVPR(2020)"

PoolNet"A simple pooling-based design for real-time salient object detection.CVPR(2019)"

Aforementioned CoSOD models should be downloaded to weights/.

Then run evaluate.ipynb for evaluation.

Citation

To be updated.