Abstract
We introduce a novel cross-reference image quality assessment method thateffectively fills the gap in the image assessment landscape, complementing thearray of established evaluation schemes -- ranging from full-reference metricslike SSIM, no-reference metrics such as NIQE, to general-reference metricsincluding FID, and Multi-modal-reference metrics, e.g., CLIPScore. Utilising aneural network with the cross-attention mechanism and a unique data collectionpipeline from NVS optimisation, our method enables accurate image qualityassessment without requiring ground truth references. By comparing a queryimage against multiple views of the same scene, our method addresses thelimitations of existing metrics in novel view synthesis (NVS) and similar taskswhere direct reference images are unavailable. Experimental results show thatour method is closely correlated to the full-reference metric SSIM, while notrequiring ground truth references.