The Open World of Micro-Videos

We examine an increasingly prevalent form of media known as micro-videos, time-constrained (typically 5-10 second) video clips commonly used on social networking sites such as Instagram and Vine. Micro-videos can be interpreted as the visual analog of a character-limited microblogs or “tweets”. An estimated 12 million micro-videos are posted to Twitter each day. The number of micro-videos produced surpasses the total inventory of YouTube every 3 months. From an applied perspective, this flood of visual data is increasingly important and has unique characteristics that are not addressed by existing computer vision methodologies and benchmarks. We further argue that the microvideo format offers unique opportunities for basic research in building systems that address lifelong learning in open-world visual domains. In this work, we introduce and analyze a micro-video dataset with 260 thousand microvideos labeled with 58 thousand tags. Example frames from the dataset are shown in Figure 1.