In this paper, we propose a novel method for quality assessment of neural network-based view synthesis (NVS), called NVS-SQA. Existing NVS quality assessment methods rely on full-reference methods such as PSNR, SSIM, and LPIPS, which have limitations due to the absence of dense reference views and the difficulty of obtaining human perceptual labels. To solve these problems, NVS-SQA proposes a method to learn reference-free quality representations through self-supervised learning without human labels. Instead of following the assumptions of existing self-supervised learning, we use heuristic cues and quality scores as learning objectives, and improve the learning efficiency through a special contrastive pair preparation process. Experimental results show that NVS-SQA significantly outperforms 17 existing reference-free methods and 16 full-reference methods.
