Domain-Separation-Graph-Neural-Networks-for-Saliency-Object-Ranking

Official implementation of the CVPR 2024 paper Domain Separation Graph Neural Networks for Saliency Object Ranking.

Installation

Our code is primarily based on MMDetection. Please refer to the MMDetection Installation for installation instructions.

Dataset

Download the ASSR Dataset and IRSR Dataset.

Training

ASSR Dataset

For resnet-50 backbone model:

bash ./tools/dist_train.sh configs/mask2former_sor/mask2former_sor_r50_assr.py num_gpus --load-from pertrained_model_path

For swin-L backbone model:

bash ./tools/dist_train.sh configs/mask2former_sor/mask2former_sor_swin-l-int21k_assr.py num_gpus --load-from pertrained_model_path

IRSR Dataset

For resnet-50 backbone model:

bash ./tools/dist_train.sh configs/mask2former_sor/mask2former_sor_r50_irsr.py num_gpus --load-from pertrained_model_path

For swin-L backbone model:

bash ./tools/dist_train.sh configs/mask2former_sor/mask2former_sor_swin-l-int21k_irsr.py num_gpus --load-from pertrained_model_path

Testing

ASSR Dataset

For resnet-50 backbone model:

bash ./tools/dist_test.sh configs/mask2former_sor/mask2former_sor_r50_assr.py model_path 1 --eval mae

For swin-L backbone model:

bash ./tools/dist_test.sh configs/mask2former_sor/mask2former_sor_swin-l-int21k_assr.py model_path 1 --eval mae

IRSR Dataset

For resnet-50 backbone model:

bash ./tools/dist_test.sh configs/mask2former_sor/mask2former_sor_r50_irsr.py model_path 1 --eval mae

For swin-L backbone model:

bash ./tools/dist_test.sh configs/mask2former_sor/mask2former_sor_swin-l-int21k_irsr.py model_path 1 --eval mae

Pretrained Models

Model	Dataset	Download
Pertrained-Res50	COCO	mask2former_r50_lsj_8x2_50e_coco
Pertrained-SwinL	COCO	mask2former_swin-l-p4-w12-384-in21k_lsj_16x1_100e_coco-panoptic

Results

Model	Dataset	SA-SOR	Download
DSGNN-Res50	ASSR	0.716	model (3qm5) | visualization results (d8m1)
DSGNN-SwinL	ASSR	0.761	model (1pjw) | visualization results (9esz)
DSGNN-Res50	IRSR	0.569	model (mfdh)
DSGNN-SwinL	IRSR	0.607	model (sq1r)

Citation

@InProceedings{Wu_2024_CVPR,
    author    = {Wu, Zijian and Lu, Jun and Han, Jing and Bai, Lianfa and Zhang, Yi and Zhao, Zhuang and Song, Siyang},
    title     = {Domain Separation Graph Neural Networks for Saliency Object Ranking},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {3964-3974}
}