NeRF-NQA: No-Reference Quality Assessment for Scenes Generated by NeRF and Neural View Synthesis Methods.

Area of Science:

• Computer Vision
• Computer Graphics
• Image Processing

Background:

• Neural View Synthesis (NVS) generates high-fidelity scenes from sparse views.
• Existing quality assessment metrics (PSNR, SSIM, LPIPS) are inadequate for NVS-synthesized dense views.
• Lack of dense ground truth limits full-reference assessment in NVS.

Purpose of the Study:

• To propose the first no-reference quality assessment method for NVS and NeRF variants.
• To address limitations of existing metrics in evaluating perceptual quality of synthesized dense views.
• To develop a method capable of assessing both spatial and angular quality aspects.

Main Methods:

• NeRF-NQA employs a joint quality assessment strategy.
• Utilizes a viewwise approach for spatial quality and inter-view consistency.
• Incorporates a pointwise approach for angular quality of scene surface points.

Main Results:

• NeRF-NQA significantly outperforms 23 mainstream visual quality assessment methods.
• Demonstrates substantial superiority in assessing NVS-synthesized scenes without references.
• Achieves effective evaluation of perceptual quality, including spatial and angular attributes.

Conclusions:

• NeRF-NQA is the first no-reference quality assessment for NVS and NeRF variants.
• The proposed method overcomes limitations of traditional metrics for dense view synthesis.
• NeRF-NQA provides a robust solution for evaluating the perceptual quality of synthesized scenes.