OpenRR-5k

The OpenRR-5k dataset is a large-scale benchmark for single-image reflection removal (SIRR) in the wild, introduced as part of the NTIRE 2026 Challenge on Single Image Reflection Removal in the Wild: Datasets, Results, and Methods.

The dataset consists of real-world images covering a variety of reflection scenarios and intensities.

GitHub Repository: caijie0620/OpenRR-5k

🥇🥈🥉 NTIRE 2026 SIRR Challenge Award Winners

Prize	Team	PSNR ↑	SSIM ↑	LPIPS ↓	DISTS ↓	NIQE ↓	Subjective ↑
🥇 1st	RRay	36.17	0.9758	0.0235	0.0135	3.7375	4.45
🥈 2nd	Xreflect Master	36.05	0.9776	0.0210	0.0127	3.7648	4.31
🥈 2nd	AIIALab	35.38	0.9750	0.0231	0.0155	3.7737	4.23
🥉 3rd	VIP Lab	34.69	0.9766	0.0231	0.0148	3.7218	3.85
🥉 3rd	YuFans	34.91	0.9738	0.0257	0.0159	3.7783	3.59
🥉 3rd	KLETech-CEVI	34.54	0.9748	0.0242	0.0150	3.7566	3.57

Dataset Structure

The dataset consists of the following components:

• train_5000.zip: contains 5,000 paired input images and corresponding ground truth (GT) images.

• val_300_blended.zip: contains 300 validation input images.

• val_300_transmission.zip: contains 300 validation ground truth images.

• test_100_blended.zip: contains 100 test input images (without ground truth).

• NTIRE2026_SIRR_TopTeam_Results.zip: Visual Results for Top-6 Teams on Val and Test Sets.

Citation

If you find this dataset helpful in your research, please cite the following work:

@article{cai2026ntire,
  title={NTIRE 2026 Challenge on Single Image Reflection Removal in the Wild: Datasets, Results, and Methods},
  author={Cai, Jie and Yang, Kangning and Li, Zhiyuan and Vasluianu, Florin-Alexandru and Timofte, Radu and Li, Jinlong and Shen, Jinglin and Meng, Zibo and Cao, Junyan and Zhao, Lu and others},
  journal={arXiv preprint arXiv:2604.10321},
  year={2026}
}

@inproceedings{cai2025openrr,
  title={Openrr-5k: A large-scale benchmark for reflection removal in the wild},
  author={Cai, Jie and Yang, Kangning and Ouyang, Ling and Fu, Lan and Ding, Jiaming and Shen, Jinglin and Meng, Zibo},
  booktitle={2025 IEEE 8th International Conference on Multimedia Information Processing and Retrieval (MIPR)},
  pages={14--19},
  year={2025},
  organization={IEEE}
}