[NeurIPS, 2025] “Image Stitching in Adverse Condition: A Bidirectional-Consistency Learning Framework and Benchmark”. Zengxi Zhang, Junchen Ge, Zhiying Jiang, Miao Zhang, Jinyuan Liu*.
[2025-10-12] The testing code for ACDIS is available
[2025-10-12] The proposed ASIS dataset is available
git clone https://github.com/ZengxiZhang/ACDIS.git
cd ./ACDIS
The demo has been tested on CUDA version of 12.1.
cd ./ACDIS
conda create -n acdis python==3.9
conda activate acdis
sh install.sh
Download the pre-trained enhancement model and put it in 1_alignment/enhancement/snapshot/:
Download the pre-trained alignment model and put it in 1_alignment/snapshot/:
Download the pre-trained composition model and put it in 2_composition/model/:
take haze environment for example:
cd 1_alignmnent
python inference_haze.py
cd ../2_composition
python inference.py
we released an Adverse Scene Image Stitching Dataset~(ASIS) that integrates in low light, haze and underwater environments. A visual representation of the dataset is shown in (c) of above figure. ASIS contains 2,250 pairs of images, including 750 images each of low-light, underwater, and haze environments, covering 17 scenes such as caves, wrecks and fields.
It is worth mentioning that some of the sources of these data come from the Internet and some come from independent photography. The captured images are far from planar structures, which ensures the disparity diversity of the proposed ASIS dataset.
We further generate the homography reference of ASIS, the generation process is shown in above figure. (The reference document contains the offset of the endpoint)
If you find our work useful in your research, please cite our paper:
@article{zhang2026acdis,
title={Image Stitching in Adverse Condition: A Bidirectional-Consistency Learning Framework and Benchmark},
author={Zhang, Zengxi and Ge, JunChen and Jiang, Zhiying and Ma, Long and Liu, Jinyuan and Fan, Xin and Liu, Risheng},
journal={Advances in Neural Information Processing Systems},
year={2025},
}