Linking Traditional TV Advertising to Internet Advertising network bandwidth (or progressive downloading), for non-intrusively embedding ads into images. The relevance between ads and images can be measured based on user-generated tags or im- age annotation. However, contextual advertising solutions from leaders like Google AdSense and BritePic (britepic.com) rely heavily on text-based models, rather than on an automated method that quickly indexes, filers, and classifies images and videos to place targeted ads alongside rich media. Visual content is basically used to locate the less intrusive regions for placing ads (Li, Wang, Yan, & Xu, 2005) (Liu, Jiang, Huang, & Xu, 2008). (Mei T., Hua, Yang, & Li, 2007) tried to detect a set of candidate video ad insertion points by locat- ing the positions with higher content discontinty and lower attrativness within an online video. Innovative research has tried to associate rel- evant ads by matching characteristic images like logos (Liao, Chen, & Hsu, 2008) or detecting high- level semantic concepts (Mei T., Hua, Yang, & Li, 2007) (Wang, Fang, & Lu, 2008). For example, "car" in a road scene is clearly related to car ads. A cell-phone video ad can be triggered when the user's mouse moves on the "phone" in a video scene. In a generic sense, the scalability of such advertising systems yet relies much on the success of exiting text based online ads, as automatic im- age and video annotation has proved to be a very challenging research problem (Wang, Zhang, Jing, & Ma, 2006). Most existing approaches attempt to learn models from a small-scale training set. Unfortunately, those models cannot be applied to large-scale web images. An alternative approach is to impose domain knowledge on visual content analysis (Gong, Lim, & Chua, 1995) (Duan, Xu, Chua, Tian, & Xu, 2003) (Li, Wang, Yan, & Xu, 2005) when scanning through images and videos for appropriateness as well as contextual relevance. For example, (Wan & Xu, 2006) used domain features to compute viewer relevance in soccer and tennis videos, where less intrusive regions like defocused background areas in a close-up shot are segmented for advertising content placement in sports highlights. The technology of behavioral targeting deliv- ers relevant ads by tracking user behavior like past search, browsing and clicking. For example, search profiling allows marketers to serve users ads using query history. Targeted ads can associ- ate categories of interest, say sports, gardens, cars, pets, with user browser, based on the types of sites users visit and the pages they view. MSN search query logs were found to be useful in finding advertising keywords on web pages (Yih, Goodman, & Carvalho, 2006). At the heart of behavioral targeting is a learning-based investiga- tion of consumer behaviors over time (Agichtein, Brill, & Dumais, 2006) (Agichtein, Brill, Dumais, & Ragno, 2006). However, behavioral targeting does have a few problems. For the first, it may be effective to target ads to a well motivated audi- ence, whereas lots of potential prospects may not reflect a user' real interest through past behavior. Moreover, some folks never learn about new products unless they are advertised, and advertis- ers could be missing lots of sales opportunities. In a sense, behavioral targeting has the effect of reducing reach. The second problem is with the privacy issues of behavioral targeting. Popular companies like Google have been struggling to provide transparency, choice, and control around privacy safety (Google Blog, 2009). TELEVISION ADVERTISING A television advertisement ­ often just commercial or TV ad ­ is span of television programming produced and paid for by an organization that conveys a message (Wikipedia). The vast major- ity of television advertisements today consist of brief adverting spots, ranging in length from a few second to a few minutes. The average cost of producing a 30-second national TV ad is over $300,000. A 30-second spot may easily run in excess of $100,000. The cost of producing a 171