A benchmark dataset for objective quality assessment of view synthesis for neural radiance field (NeRF)

Abstract

Neural Radiance Fields (NeRF) are revolutionizing diverse fields such as autonomous driving, education, and virtual reality (VR). As their applications expand, the ability to accurately evaluate the quality of NeRF-generated content becomes essential. Currently, there are only a few datasets for NeRF quality evaluation. Also, while existing quality datasets primarily utilize processed video sequences (PVS) as stimuli, real-world scenarios often involve uneven camera trajectories, underscoring the need for alternative approaches to subjective quality assessment. This study proposes a quality dataset for assessing the quality of NeRF. The dataset was generated by varying quality parameters in SOTA NeRF models to create different quality levels. A subjective experiment was conducted to obtain human opinion scores for the distorted NeRF. The subjective data were processed in accordance with International Telecommunication Union (ITU) guidelines to derive mean opinion scores (MOS. The datasets and findings not only offer insights into the performance of NeRF models but also serve as valuable resources for developing quality assessment models.