NeRF quality dataset for evaluating objective quality models

This repository hosts a dataset for Neural Radiance Field (NeRF) research. The dataset contains 102 synthesized stimuli, which were designed with non-overlapping scenes. These stimuli were generated using four state-of-the-art (SOTA) NeRF models: DVGO, InstantNGP, Plenoxels, and TensoRF, by systematically varying key quality parameters to create different quality levels. A remote subjective experiment was conducted with 86 participants to annotate the dataset following detailed guidelines. The collected opinion scores were processed to obtain the Mean Opinion Score (MOS) in accordance with the ITU-R BT.500-15 standard. The MOS scores are in a csv file named Subjective_mos.csv. This dataset can be use to facilitate the refining of image quality assessment models for NeRF.

本仓库收录了一套面向神经辐射场（Neural Radiance Field, NeRF）研究的专用数据集。该数据集包含102个合成刺激样本，所有样本均采用互不重叠的场景构建。这些刺激样本由四款当前最先进（state-of-the-art, SOTA）的NeRF模型生成，分别为DVGO、InstantNGP、Plenoxels与TensoRF；生成过程中通过系统性调整关键质量参数，以生成不同质量等级的样本。研究团队开展了远程主观实验，邀请86名受试者按照详细指导规范对数据集进行标注。随后，依据ITU-R BT.500-15标准对收集得到的主观评分进行处理，计算得到平均意见得分（Mean Opinion Score, MOS）。该平均意见得分存储于名为Subjective_mos.csv的CSV文件中。本数据集可用于推动面向NeRF的图像质量评估模型的优化与完善。