合成GS/RS数据集

本研究利用Unreal Engine 5（UE5）创建了一个合成数据集，专门用于评估不同快门机制下的行人检测模型。数据集包含40个不同的城市街道场景，每个场景在五个不同的时间点进行渲染，以模拟多样的光照条件。此外，场景中随机生成的行人群体增加了数据的多样性和真实性。数据集的创建过程涉及高帧率的全局快门（GS）图像生成，并通过模拟滚动快门（RS）效应来合成RS数据集。该数据集主要应用于机器学习模型在行人检测中的性能评估，特别是在考虑RS效应时，旨在优化无ISP的机器学习管道，提高检测精度。

This study utilized Unreal Engine 5 (UE5) to create a synthetic dataset specifically designed for evaluating pedestrian detection models under different shutter mechanisms. The dataset includes 40 distinct urban street scenes, each rendered at five different time points to simulate diverse lighting conditions. Additionally, randomly generated pedestrian crowds within the scenes enhance the data's diversity and realism. The dataset creation process involves generating high-frame-rate global shutter (GS) images, and synthesizing RS datasets by simulating the rolling shutter effect. This dataset is primarily used for performance evaluation of machine learning models in pedestrian detection, particularly when accounting for the RS effect, with the goal of optimizing ISP-free machine learning pipelines and improving detection accuracy.