About The Forum
Calendar 2009 Poster Sessions : Linking Images Into Useful Webs
Abstract:
The widespread availability of digital cameras, coupled with ubiquitous Internet access and decreasing storage costs, have made it easy to capture and share large amounts of image data. Photo sharing sites and large-scale image collection projects like Google Street View have generated massive image collections where many images are related in the sense that they contain similar objects or view parts of the same scene. Such relationships between images can conceptually ``connect'' the collection in interesting ways. In this paper, we address the problem of how to discover, represent, and utilize relationships among images with similar parts. Our approach starts by extracting corresponding patches from pairs of images by finding geometrically consistent sets of local feature correspondences. It then forms links between these patches --- both across images and within the same image. These links define a graph structure on the image patches that we call an \emph{Image Web}, to which we apply graph traversal and link analysis techniques to discover interesting structure within the collection. We describe a pipeline for building Image Webs from collections of images and demonstrate how Image Webs can serve as a unifying framework for a diverse set of applications.
These applications include a novel way to browse a collection of images, discover relationships between a set of images, organize a community photo collection, and transfer annotations to unannotated images.
Bio:
Kyle is PhD candidate in the department of Electrical Engineering at Stanford University. His research interests include computer vision and wireless sensor networks.
The widespread availability of digital cameras, coupled with ubiquitous Internet access and decreasing storage costs, have made it easy to capture and share large amounts of image data. Photo sharing sites and large-scale image collection projects like Google Street View have generated massive image collections where many images are related in the sense that they contain similar objects or view parts of the same scene. Such relationships between images can conceptually ``connect'' the collection in interesting ways. In this paper, we address the problem of how to discover, represent, and utilize relationships among images with similar parts. Our approach starts by extracting corresponding patches from pairs of images by finding geometrically consistent sets of local feature correspondences. It then forms links between these patches --- both across images and within the same image. These links define a graph structure on the image patches that we call an \emph{Image Web}, to which we apply graph traversal and link analysis techniques to discover interesting structure within the collection. We describe a pipeline for building Image Webs from collections of images and demonstrate how Image Webs can serve as a unifying framework for a diverse set of applications.
These applications include a novel way to browse a collection of images, discover relationships between a set of images, organize a community photo collection, and transfer annotations to unannotated images.
Bio:
Kyle is PhD candidate in the department of Electrical Engineering at Stanford University. His research interests include computer vision and wireless sensor networks.