Download the pretrained model weights from OneDrive and place them in the corresponding directories:

• Model File: checkpoint/assr_swinl/model.pth
• Model File: checkpoint/irsr_swinl/model.pth

You can download the datasets from their official repositories:

• ASSR: Attention_Shift_Ranks
• IRSR: Saliency-Ranking

Alternatively, we also provide the required data on OneDrive for easy reproduction. Download and extract the data, then place it under the data/ directory.

The simplest way is to run demo.sh directly:

bash demo.sh

Inference on ASSR dataset:

python main.py --num-gpus 1 \
    --config-file configs/coco/instance-segmentation/swin/maskformer2_swin_large_IN21k_384_bs16_100ep_assr.yaml \
    EVALUATION.DATASET "assr" \
    EVALUATION.MODEL_DIR "checkpoint/assr_swinl/" \
    EVALUATION.MODEL_NAMES "('model.pth', )"

After inference is complete, run the evaluation script:

python metric.py --map output/assr_swinl/model/ResultThres

We also provide our pre-computed prediction results for direct evaluation. Download from OneDrive and evaluate using the same script above.

1. Ensure sufficient GPU memory (recommended >= 16GB)
2. For the first run, CUDA operators need to be compiled:

   cd mask2former/modeling/pixel_decoder/ops
   sh make.sh

3. If you modify the confidence threshold, the output path will change accordingly

@inproceedings{liu2025language,
  title={Language-guided salient object ranking},
  author={Liu, Fang and Liu, Yuhao and Xu, Ke and Ye, Shuquan and Hancke, Gerhard Petrus and Lau, Rynson WH},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={29803--29813},
  year={2025}
}

For any questions or issues, please contact: fawnliu2333@gmail.com