HSD (heterogeneous stereo deblurring dataset)
收藏资源简介:
HSD数据集是由釜庆国立大学和建国大学联合构建的异构立体去模糊基准数据集,旨在研究智能手机异构相机系统在硬件差异下导致的视间不对称模糊问题。该数据集包含384个场景序列,最终生成了总计3,300个图像对(训练集2,200对,测试集1,100对),数据来源于iPhone等商用智能手机同步采集的宽视角和超广角视频流,并经过设备级几何校正。数据集的构建通过多帧时序积分模拟超广角视图的曝光运动模糊,同时以宽视角作为清晰参考视图,从而在真实移动成像管道中实现对不对称模糊的受控评估。该数据集主要应用于计算机视觉和计算摄影领域,特别是异构立体图像恢复任务,为解决智能手机多相机系统中因硬件不对称性(如光学防抖缺失、光圈差异)引起的跨视图质量退化问题提供了专门的评估基准。
The HSD Dataset is a heterogeneous stereo deblurring benchmark dataset jointly constructed by Pukyong National University and Konkuk University, aiming to investigate the inter-view asymmetric blur problem caused by hardware differences in heterogeneous camera systems of smartphones. It contains 384 scene sequences, ultimately generating a total of 3,300 image pairs, with 2,200 pairs for the training set and 1,100 pairs for the test set. The data is sourced from synchronously captured wide-angle and ultra-wide-angle video streams using commercial smartphones such as iPhone, and has undergone device-level geometric calibration. The dataset is built by leveraging multi-frame temporal integration to simulate exposure-induced motion blur for ultra-wide-angle views, while taking wide-angle views as sharp reference views, thereby enabling controlled evaluation of asymmetric blur in real-world mobile imaging pipelines. This dataset is mainly applied in the fields of computer vision and computational photography, particularly for heterogeneous stereo image restoration tasks, providing a dedicated evaluation benchmark to address cross-view quality degradation issues caused by hardware asymmetries in smartphone multi-camera systems, such as lack of optical image stabilization and aperture differences.
数据集概述
- 数据集名称: Heterogeneous Stereo Deblurring (HSD) 数据集
- 论文标题: A Benchmark for Heterogeneous Stereo Deblurring with Physically- and Epipolar-constrained Cross Attention
- 会议: ECCV 2026
- 许可证: 数据集采用 CC BY 4.0 许可证
数据集简介
HSD 数据集专为解决异构立体去模糊问题而构建,针对现代智能手机中不同摄像头模块导致的非对称运动模糊伪影。数据集来源于真实智能手机的立体捕获,通过多帧集成技术构建而成。
数据集结构
Heterogeneous-Stereo-Deblurring/ ├── train/ │ ├── input/ │ ├── target/ │ └── guide/ └── test/ ├── input/ ├── target/ └── guide/
- train/: 训练集,包含输入图像(input)、目标清晰图像(target)和引导图像(guide)。
- test/: 测试集,结构与训练集相同。
下载地址
实验结果
在 HSD 数据集上,使用 PECA 模块增强的模型在立体输入下取得了更优的性能:
| 骨干网络 | 变体 | 输入类型 | PSNR (dB) | SSIM |
|---|---|---|---|---|
| XYDeblur | 基线 | 单目 | 30.76 | 0.9444 |
| XYDeblur | PECA | 立体 | 32.22 | 0.9620 |
| Restormer | 基线 | 单目 | 30.94 | 0.9444 |
| Restormer | PECA | 立体 | 32.47 | 0.9636 |
| NAFNet | 基线 | 单目 | 32.16 | 0.9579 |
| NAFNet | PECA | 立体 | 32.92 | 0.9669 |
引用信息
请使用以下 BibTeX 引用该工作:
@inproceedings{shin2026PECA, title={A Benchmark for Heterogeneous Stereo Deblurring with Physically- and Epipolar-constrained Cross Attention}, author={Shin, Hoju and Kim, Jiah and Kim, Seung-Wook and Ji, Seowon}, booktitle={ECCV}, year={2026} }

- 1A Benchmark for Heterogeneous Stereo Deblurring with Physically- and Epipolar-constrained Cross Attention釜庆国立大学·智能机器人工程系;建国大学·计算机科学与工程系 · 2026年




