Pooling is the step that collapses a full-reference metric's many per-frame scores into the single number you report for a clip. A ten-second clip at 30 frames per second produces 300 scores, and pooling combines them the way a teacher turns many marked assignments into one course grade. The catch is that the per-frame scores are fixed but the headline number depends entirely on the rule you choose, so the reported score can move without a single pixel changing. The default in nearly every tool is the arithmetic mean, which weights every frame equally and therefore hides a short stretch of bad quality inside a long good one. Better choices surface the worst frames: the harmonic mean tilts toward low values, percentile pooling reports the worst 5 or 10 percent, and the minimum reports the single worst frame. Pooling happens twice, once across pixels inside a frame (spatial) and once across frames over time (temporal). Always state which pooling method you used.
The handbook · 7 volumes
Fora Soft Video Engineering Handbook
Pooling turns per-frame scores into one number, and the method decides whether bad frames show up or get averaged away. We wrote seven volumes covering the whole video stack, and for pooling start with Volume 1: per-frame metrics, VMAF, PSNR and SSIM and how scores are reported, with Volume 3 on QoE. Written by Nikolay Sapunov, CEO at Fora Soft.
Not sure which pooling method your quality reports should use?
Send us per-frame data from your encodes: we compare mean, harmonic and percentile pooling and recommend the one for your gates and reports. 250+ video products shipped since 2005.
