CueNet: a context discovery framework to tag personal photos

An image recognition problem is typically formulated as tagging a given set of images with labels from a predefined set. Context-aware approaches in problems like face recognition have utilized information about a user and the people she knows through different social networks. Traditionally, this context is statically linked to all of the available data. In this work, we propose a technique to dynamically discover which subset of all the available data is relevant context for the given recognition problem. In this dissertation, we propose the CueNet framework, to discover candidate labels for the person identification problem in personal photos. We describe our context model, and how it allows heterogeneous data sources to contribute useful context for the identification problem. We design algorithms to extract contextual information from these sources to discover a subset of candidates who could potentially appear in personal photos. Our early experiments show that CueNet is capable of removing upto 99% of irrelevant candidates, and was able to correctly tag 80% of frontal faces.