GeoLookbook: Modeling Worldwide Human Visual Appearance

Every day millions of images, with scenes of daily life from around the world, are uploaded and made publicly to photo sharing sites. This imagery is a rich source of information about the world and how it varies over time and space. We focus on extracting information about people: their facial appearance and what clothes they wear. While there is significant work in developing automated techniques to extract biometric features from individual images, little work has attempted to model the relationship between appearance and geographic location. This type of knowledge has been restricted to human experts; we propose to learn computational models from social media imagery and make it usable by experts and novice users for a wide variety of applications. Performing this analysis poses a number of technical challenges: collecting, archiving and processing hundreds of millions of images; developing robust low-level pose normalization methods; combining information from millions of images to develop a global model; and visualizing the results of the analysis. We will a robust, large-scale image processing system and a collection of web-based visualize systems that support specific applications.