The complement starts off by collecting suit info from both OIS and a phone’s gyroscope, creation certain it’s in “perfect” sync with a image. But it’s what happens subsequent that matters most: Google uses a “lookahead” filtering algorithm that pushes picture frames into a deferred reserve and uses appurtenance training to envision where you’re expected to pierce a phone next. This corrects for a wider operation of transformation than OIS alone, and can negate common video quirks like wobbling, rolling shiver (the exaggeration outcome where tools of a support seem to loiter behind) or concentration hunting. The algorithmic process even introduces practical suit to facade furious variations in sharpness when we pierce a phone quickly.
This isn’t to contend that Google’s proceed is flawless. As others have noted, a Pixel 2 can stand a support in astonishing ways and fuzz low light footage some-more than it should. On a balance, though, this shows only how most AI-related record can assistance with video. It can erase standard errors that EIS or OIS competence not locate by themselves, and produces footage so well-spoken it can demeanour like it was prisoner with the assistance of a gimbal.