A no-reference metric, also called blind, judges the quality of the impaired video alone, with nothing to compare against, the way a person can glance at a photo and say it looks compressed without ever seeing the original. It is the hard case but also the most common one in the real world: live broadcast, video calls, surveillance, and user-generated clips have no pristine master, because the camera feed already is the only version that exists, which makes no-reference the only option there. Such metrics come in three styles: natural-scene-statistics models like NIQE and BRISQUE, learned deep networks such as a no-reference VMAF or Google UVQ, and bitstream or parametric models like ITU-T P.1203 and P.1204.3 that read encoded metadata. The catch is accuracy: with no original to anchor against, no-reference is the hardest family to trust, varies widely by content, and can be fooled, so its scores are best read as trend signals, not ground truth. It sits opposite full-reference, with reduced-reference in between.
The handbook · 7 volumes
Fora Soft Video Engineering Handbook
A no-reference metric judges quality from the delivered video alone, which is the only option for live streams and user uploads. We wrote seven volumes covering the whole video stack, and for no-reference metrics start with Volume 1: perceptual and no-reference metrics, with Volume 3 on QoE monitoring and Volume 6 on models and inference. Written by Nikolay Sapunov, CEO at Fora Soft.
Need quality scores on live or UGC video where no source exists?
Tell us your content and scale: we pick and benchmark a no-reference metric on your streams and give you a monitoring plan with costs. 250+ video products shipped since 2005.
