Article ID: 2024IMP0003
This paper presents a novel reduced-reference image quality assessment (RR IQA) method from monocular dynamic scene images for neural radiance fields (NeRF). Despite recent advancement in NeRF, evaluating the performance of NeRF models remains challenging due to the difficulty associated with obtaining ground truth viewpoint images for dynamic scenes. Collecting such ground truth images for NeRF model evaluation typically requires capturing the target scene from multiple synchronized cameras, which is labor-intensive. To address this issue, we propose a novel RR IQA metric called amplitude-dissimilarity (AMDIS), which focuses on evaluating NeRF models without requiring ground truth viewpoint images. The key idea behind AMDIS is that the differences between two near-viewpoint images are mainly absorbed in the phase components.
Thus, AMDIS evaluates NeRF models by measuring the dissimilarity between the Fourier amplitude components of the training and synthesized images. Because AMDIS only uses the training and synthesized images, the corresponding ground truth viewpoint images are not required for the evaluation. The experimental results demonstrate that the proposed AMDIS is strongly correlated with major full-reference IQA methods that directly use ground truth viewpoint images.
