2018年视频内容理解在Hulu的应用与实践.pdf

《2018年视频内容理解在Hulu的应用与实践.pdf》由会员分享，可在线阅读，更多相关《2018年视频内容理解在Hulu的应用与实践.pdf（27页珍藏版）》请在三个皮匠报告上搜索。

1、Applications and best practice:video understanding in HuluAbout Hulu and its contentsWhy video understanding in Hulu and its challenges Applications of video understanding in HuluBest practice and applications in Hulu business and production Hulu AI platform and AI automation pipeline5Best Quality C

2、ontentVideo-On-Demand and Live Broadcasting Hulu,the Premier Digital Video Company67540CONTENT PARTNERS81000+ADVERTISNG PARTNERSAs a content platform:Live channels 1000+Movies:23wTV series:55w,episodes:430wAds video:185wShort form video:200wData growsData grows24%CAGR 16-20Exabytes per monthTech bre

3、akthroughTech breakthroughIndustry trendingIndustry trendingBusiness requirementBusiness requirementWhy video understandingCisco:Global consumer internet traffic,20162021Feifei Li,ImageNet classification errors,CVPR 172016.10 releasedTechnical challenges in video understanding数据的数据的多与少多与少4.5M videos

4、/22M+usersRich scenarios with diversity and small labelled dataset数据的数据的真与假真与假Sports,news,Si-Fi,cartoon,cosmetic,技技术的术的难与易难与易Temporal sequential frame analysisSemantic gap企企业的业的买与研买与研AI platform and upgrade of tech stacks Hulu AI platform Share Model/API Share Data Share InfrastructureHulu AI automa

5、tion pipelineVideo derived metadata1.Fine grain video segmentation2.Video derived Tags3.Content generation&highlights4.Practice&applications in productionContent understanding application categoriesAntecedentsOpeningRating(burn-in)captionAds breaksoundtrack(music,sound,.)Shot boundaryBrandingEnd cre

6、dits markerBurn-in channel logoGreen:well resolvedDark:not started1.Fine grain video segmentationThumbnailEnd credits infoModelError 5s(%)Error 10s(%)Existing markers in theR database62.6484.31Our hybrid Deep CNN model 86.8692.531.1 End credits detection1.2 Burnt-in video logo detection300+logos,97.