| Recent years have seen an explosion of internet image collections. The explosion was brought about by the popularity of photo sharing sites such as Flickr and Picasa Web Albums, where millions of internet users upload their personal photos online and share these photos with the public. The wealth of images provides an excellent resource for perceiving the world through the eyes of others. However, the gigantic volume of images also poses a challenge to the consumption of this unorganized visual information.;In this dissertation, we present research on canonical view mining. Given an image collection, we leverage a combination of computer vision and data mining techniques to infer and remove images of noisy and redundant views. The remaining images, which we term canonical views, exhibit both representativeness and diversity in image content, and form a succinct visual summary of the original image collection.;The main contribution of this dissertation is the development and evaluation of a fully automatic pipeline for canonical view mining. We also demonstrate two applications of canonical views in the context of image browsing and object recognition. Finally, we analyze the scalability of the pipeline for canonical view mining and propose an approximation algorithm that effectively removes the scalability bottleneck with low impact on the resulting canonical views.;Keywords: canonical views, internet images, the wisdom of crowds, place recognition, scalable image matching, kd-tree... |
