It's easy not to think about just how much thought and computing goes into everything that Google does, but one of YouTube's latest changes reminds us of precisely that.
Every video on the site has a thumbnail that's supposed to offer you a glimpse into what you're about to watch. A bad image will discourage you from clicking. Good ones lead to more views and greater revenue. So naturally both content creators and Google would prefer to have better thumbnails.
Using deep neural networks, the YouTube team has launched an improved "thumbnailer." Every frame in a video gets evaluated by a quality model and is assigned a quality score. The highest scoring options get enhanced and rendered at different sizes and aspect ratios.
Google trained its quality model by showing it good and bad thumbnails. Postive examples tended to come from YouTubers who created custom thumbnails that were well framed and in-focus, with a particular subject usually in the center.
This training has resulted in an automatic thumbnailer that picks images humans preferred 65% of the time over pictures selected by the old system.
The behind-the-scenes magic is already at work on the site, so the next time you upload a video, give the automatically generated thumbnails an extra look.
- Google Research Blog