Product and Brand tracking over social media is very important for generating business intelligence for companies to better understand the online sentiments, interests and concerns of customers. Increasingly, users are sharing more multimedia content over the social network, and over 50% of these multimedia posts do not have meaningful text annotations. In order to track brand or event information more completely over the social media platforms, the ability to process visual content in conjunction with text analysis is very important.
This research focuses initially on product and brand tracking. The key issue in brand tracking in microblogs is not search, but the ability to gather sufficiently representative set of microblogs, filter out those irrelevant, and process the rest for various applications such as event detection and tracking. Given a brand to track with a multimedia query (a text term and an example logo), we developed a framework to tackle this problem by (See Figure 2.6): (a) performing a text-based search to find all microblogs containing the brand name; (b) finding those returned microblogs with images containing the visual logo by performing logo detection, and forming a highly relevant “seed set” of microblogs where both text and images are relevant to the brand; (c) performing extended searches from the “seed set” to increase the coverage by utilizing both social and visual context; and (d) filtering our irrelevant microblogs by performing joint text-visual fusion. Through this framework, we are able to obtain promising results with higher coverage and accuracy.
To promote this line of research, we are building a large-scale dataset with ground truth for logos, products and brands. The dataset comprises over 3 million Weibo microblogs with over 1 million images and cover 100 logos, 100 brands and 300 products. We will release our dataset towards the end of 2013.
Figure 2.6. The architecture for brand tracking in microblogs
 Fangling Wang, Yue Gao, Tat-Seng Chua: Brand Tracking in Microblogs. Internal Report.
 Guangda Li, Meng Wang, Zheng Lu, Richang Hong, Tat-Seng Chua: In-Video Product Annotation with Web Information Mining. TOMCCAP 8(4): 55 (2012)