Using the Internet for object image retrieval and object image classification

Internet;Object image retrieval;Object image classification

Wang, Gang ; Forsyth, David A. ; Forsyth ; David A.

关键词: Internet;    Object image retrieval;    Object image classification;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/14623/Wang_Gang.pdf?sequence=2&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

The Internet has become the largest repository for numerous resources, a big portionof which are images and related multimedia content such as text and videos. Thiscontent is valuable for many computer vision tasks. In this thesis, two case studies areconducted to show how to leverage information from the Internet for two importantcomputer vision tasks: object image retrieval and object image classification.Case study 1 is on object image retrieval. With specified object class labels, weaim to retrieve relevant images found on web pages using an analysis of text around theimage and of image appearance. For this task, we exploit established online knowledgeresources (Wikipedia pages for text; Flickr and Caltech data sets for images). Theseresources provide rich text and object appearance information. We describe results ontwo data sets. The first is Berg’s collection of 10 animal categories; on this data set, wesignificantly outperform previous approaches. In addition, we have collected 5 morecategories, and experimental results also show the effectiveness of our approach on thisnew data set.Case study 2 is on object image classification. We introduce a text-based imagefeature and demonstrate that it consistently improves performance on hard object classificationproblems. The feature is built using an auxiliary dataset of images annotatedwith tags, downloaded from the Internet. We do not inspect or correct the tags andexpect that they are noisy. We obtain the text features of an unannotated image fromthe tags of its k-nearest neighbors in this auxiliary collection. A visual classifier presentedwith an object viewed under novel circumstances (say, a new viewing direction) must rely on its visual examples. Our text feature may not change, because the auxiliarydataset likely contains a similar picture. While the tags associated with imagesare noisy, they are more stable when appearance changes. We test the performance ofthis feature using PASCAL VOC 2006 and 2007 datasets. Our feature performs well;it consistently improves the performance of visual object classifiers, and is particularlyeffective when the training dataset is small.