This project provides the code and results for 'IPDiff: Diffusion-driven ORSI Salient Object Detection with Information Reconstruction and Multi-Prior Guidance', IJCV, 2026. IJCV | pdf | arxiv (Code is coming soon)
python 3.9 + pytorch 2.1.1 + cuda 12.1
To install requirements:
pip install -r requirements.txt
For more details, please refer to CamoDiffusion.
We provide saliency maps of our IPDiff (code: IPDI) on three datasets.
We also provide saliency maps of all compared methods (code: IPDI) on three datasets.
-
Use data_aug.m for data augmentation, and put the augmented data under './ORSI_dataset_aug'.
-
Download the pretrained backbone pvt_v2_b4_m.pth (code: IPDI), and put it under './pretrained_weights/models--Anonymity--pvt_pretrained/snapshots'.
-
Set your dataset paths in the configuration file under './config/dataset_352x352.yaml' and some training hyperparameters under './configs/default.yaml'.
-
Run train_ORSSD.py/train_EORSSD.py/train_ors4199.py.
Note: Our main model is under './model/net.py'
-
Download the pre-trained models on ORSSD (code: IPDI), EORSSD (code: IPDI), and ORSI-4199 (code: IPDI), and put them in './models'.
-
Modify paths of pre-trained models and datasets under './sample.py'.
-
Run sample.py.
You can use the evaluation tool (MATLAB version) to evaluate the above saliency maps.
This work is built upon Camodiffusion.
@ARTICLE{Li_2026_IPDiff,
author = {Gongyang Li and Zhen Bai and Runmin Cong and Dan Zeng and Weisi Lin and Xiao-Ping Zhang},
title = {{IPDiff}: Diffusion-driven ORSI Salient Object Detection with Information Reconstruction and Multi-Prior Guidance},
journal = {International Journal of Computer Vision},
volume= {134},
number = {7},
pages = {341},
year = {2026},
doi = {10.1007/s11263-026-02931-y},
}
If you encounter any problems with the code, want to report bugs, etc.
Please contact me at lllmiemie@163.com or ligongyang@shu.edu.cn.