An Evaluation of Quality Metrics for Neural Radiance Field
Chibuike Onuoha, Jean Atsumi Flaherty, Shihao Luo, Trương Thu Hương, Truong Cong Thang · 2023
As Neural Radiance Field (NeRF) models unfold, the evaluation of perceptual quality becomes paramount for ensuring good quality of experience for users. This is especially important in new services like VR/AR and e-learning. In this paper, we evaluate the performance of image and video quality metrics on NeRF-synthesized contents. With a keen focus on correlating a NeRF model with human perception, we conduct a subjective study by gathering crowd opinions scores which are used as the ground truth for evaluation. The results show that deep learning based models like DISTS and CompressVQA provide the best metrics, while conventional metrics like PSNR, SSIM, and MS-SSIM have rather low performances.