MPII-Test
收藏资源简介:
MPII-Test数据集是由上海交通大学和vivo移动通信有限公司的研究人员创建的,旨在为人体图像恢复领域提供一个新的基准。该数据集包含5427张真实世界中的退化人体图像,其中许多图像具有丰富的运动模糊模式。这些图像是通过模拟人体运动模糊和通用噪声的共存情况生成的,以帮助模型学习更真实的退化场景,并在实际应用中提高其泛化能力。数据集的创建过程包括人体分割模型的应用、运动模糊模拟模块的集成以及通用两阶段退化流程。MPII-Test数据集旨在解决现有退化流程中人体运动模糊不足的问题,并提高模型在包含人体运动模糊和其他退化类型的人体图像恢复任务中的性能。
The MPII-Test dataset was created by researchers from Shanghai Jiao Tong University and Vivo Mobile Communications Co., Ltd., aiming to provide a new benchmark for the field of human image restoration. This dataset contains 5,427 real-world degraded human images, many of which exhibit rich motion blur patterns. These images are generated by simulating the coexistence of human motion blur and general noise, to help models learn more realistic degradation scenarios and improve their generalization ability in practical applications. The creation process of the dataset includes the application of human segmentation models, the integration of motion blur simulation modules, and a general two-stage degradation pipeline. The MPII-Test dataset is designed to address the insufficient human motion blur issue in existing degradation pipelines, and enhance the performance of models on human image restoration tasks involving both human motion blur and other degradation types.
HAODiff 数据集概述
数据集基本信息
- 名称: HAODiff: Human-Aware One-Step Diffusion via Dual-Prompt Guidance
- 发布日期: 2025-05-27
- 研究领域: 人体图像恢复(处理运动模糊和通用噪声)
- 论文链接: https://arxiv.org/abs/2505.19742
- 补充材料: https://github.com/gobunu/HAODiff/releases/download/Paper/supp.pdf
核心贡献
- 提出一种新型单步扩散模型(HAODiff),用于人体图像恢复
- 设计三重分支双提示引导(DPG)机制
- 引入MPII-Test基准测试集(包含混合噪声和人体运动模糊案例)
关键技术
- 退化管道: 模拟人体运动模糊(HMB)与通用噪声的共存
- 训练目标:
- 高质量图像
- 残差噪声(LQ-HQ)
- HMB分割掩码
- 双提示对生成: 在单扩散步骤中实现分类器无关引导(CFG)
评估数据集
- PERSONA-Val(合成数据集)
- PERSONA-Test(真实世界数据集)
- MPII-Test(新引入的基准测试集)
性能表现
- 在定量指标和视觉质量上均超越现有SOTA方法
- 主要评估结果见论文中的Table 1(PERSONA-Val)和Table 2(PERSONA-Test/MPII-Test)
可视化结果
- 包含合成数据集(Figure 5)和真实数据集(Figure 6)的对比
- 补充材料中提供更多纹理(Figure 4)和挑战任务(Figure 11-12)的对比
当前状态
- [ ] 代码和预训练模型待发布
- [x] 论文结果已公开
引用格式
bibtex @article{gong2025haodiff, title={{HAODiff: Human-Aware One-Step Diffusion via Dual-Prompt Guidance}}, author={Gong, Jue and Yang, Tingyu and Wang, Jingkai and Chen, Zheng and Liu, Xing and Gu, Hong and Liu, Yutong and Zhang, Yulun and Yang, Xiaokang}, journal={arXiv preprint 2505.19742}, year={2025} }




