As such a message as that, that there should be for YouTube videos from now more meaningful thumbnails, not worth big headline - even if, as Google says, the thumbnails are the first thing that potential viewers see from the clip, and a interesting start screen rather encouraged to click the Play button. Likewise, long been known that Google uses neural networks to more "intelligent", ie to implement lernfaehige automations; at all, many solutions are based in the area of u200bu200bobject recognition and computer vision on such solutions. And yet if Google proudly announced developed a new Thumbnailer and this by using Deep Neural Networks (dnns) to have taught what good thumbnails differs from bad, one can in a quiet moment alreadysome start to wonder. Initially thought to be trying to force sparrows and guns, because to put so much effort on automatically generated Vorschauansichen, seems in view of the fact that everyone can choose a matching frame from his video by hand, slightly exaggerated. On the other hand, the great mass of all Uploader will not do this, of course, and after uploaded but not played videos for Google are useless, the step for the Group could possibly be worth actually. pic1 The distinction u003csu003ebetweenu003c/su003e good and evil bad have the neural networks trained by the analysis of preview images which have been selected by humans. These are - unlike randomly selected film stills - mostly sharp, kadriert carefully and usually show especiallya main character or the principal motive of the film. If it succeeds in new Thumbnailer actually identify in each clip such a picture, that would be quite handy in itself. Or maybe not, because it is more difficult when fast scrolling, presumably to distinguish Featured Videos estimated crudely made shots because everything looks the same somehow. We therefore propose to let the dnns additionally was calculated if a sharp and well composed preview image is even statistically representative of the style of each clip, and - if not - to select a still of low quality. If already quantified, then right ... u003cimg src=
more infos at bei googleresearch.blogspot.it;)
