DS-IRP

为了研究IRP，我们首先建立了一个名为动态场景-IRP (ds-irp) 的数据集。Ds-irp的收集主要包括两个阶段。在第一阶段，我们在各种相机设置下收集了2,500个动态场景和每个场景中的11个图像。由于在动态场景中，通过在现实中拍摄来获取地面真实图像是不切实际的，而这在阶段2中是必需的，因此我们选择在动态成像形成后模拟图像，最终总共有27,500个图像。在第二阶段中，我们对从阶段1收集的每个图像重新训练并测试了4种代表性的恢复算法。然后，通过参考地面真实图像来计算恢复的图像质量来获取IRP标签。To investigate IRP, we first constructed a dataset named Dynamic Scene-IRP (ds-irp). The data collection process of ds-irp primarily comprises two stages. In the first stage, we collected 2,500 dynamic scenes, with 11 images captured for each scene under various camera configurations. Since acquiring ground-truth images through in-situ photography is impractical for dynamic scenes—a requirement for Stage 2—we elected to simulate the images post dynamic imaging formation, ultimately resulting in a total of 27,500 images. In the second stage, we retrained and tested four representative image restoration algorithms on each of the images collected from Stage 1. Subsequently, we derived the IRP labels by evaluating the quality of the restored images against the ground-truth images.

搜集汇总

背景与挑战

背景概述

DS-IRP是一个为研究动态场景中图像恢复潜力（IRP）而构建的数据集，包含27,500张图像，通过两个阶段收集：首先在多种相机设置下采集2,500个动态场景，然后模拟图像并用算法生成IRP标签。该数据集由阿德莱德大学和西北工业大学于2022年发布。