Finding Celebrities in Billions of Web Images 云飞 2012-12-13 Overview • 一、label an input image with a list of celebrities. • 二、the celebrity names are assigned to the faces by label propagation on a facial similarity graph. Overview • 本文的优点: • 1、the proposed image annotation system is capable of labeling names to general web images. • 2、our name assignment algorithm does not impose any assumption on the facial feature distribution. • 3、not only visual cues are used. Overview • 1. determine, by identifying celebrity names from surrounding text. • 2. given a set of names, assign the names to the faces in the input image. Overview • A. Image Annotation System • 1) construct a vocabulary; • 2) discover all webpages hosting its near-duplicates; • 3) use the vocabulary to filter the surrounding text. – Advances: • 1)effective; • 2)remove noise. – Annotated images: • 1)SFSN • 2)SFMN • 3)MF Overview • B. Multimodal Name Assignment • The context likelihood incorporates the information from surrounding text by using the confidence scores estimated by the image annotation system. IMAGE ANNOTATION SYSTEM • Goal: label an input image with a list of celebrities who may appear in the image. • A. Constructing a Large-Scale Celebrity Name Vocabulary • B. Discover Related Webpages by Near-Duplicate Image Retrieval • C. Annotating Images by Mining Surrounding Text of Related Webpages IMAGE ANNOTATION SYSTEM A. Constructing a Large-Scale Celebrity Name Vocabulary 1)Wikipedia 首段 信息框 标签 2)Entitycube IMAGE ANNOTATION SYSTEM • B. Discover Related Webpages by NearDuplicate Image Retrieval – divide and conquer strategy • 图片分成n×n • 降维 • 阈值化 IMAGE ANNOTATION SYSTEM • C. Annotating Images by Mining Surrounding Text of Related Webpages • 1) Type of names; • 2) Type of surrounding text; • 3) Frequency versus ratio; MULTIMODAL NAME ASSIGNMENT • • • • A. Notation B. Overview of the Assignment Model C. Label Propagation from SFSN Images p(Y|F) D. Constrain the Propagation by a Context Likelihood p(Y|T; λ) • E. Normalization by Name Prior p(Y) • F. Implementation Detail: Face Representation • A. Notation – faces in image In – denote the face labels as • B. Overview of the Assignment Model – the confidence for label • C. Label Propagation from SFSN Images p(Y|F) – how to propagate labels from SFSN images to SFMN and MF images • D. Constrain the Propagation by a Context Likelihood p(Y|T; λ) • 1) For each image-level name vk, generate a binary variable zk from p(vk |T) as defined in (3) to indicate whether vk appears in image I. • 2) If zk=1, generate mk faces of name vk in image I from p(m|z; λ) as defined in (13). • E. Normalization by Name Prior p(Y) – p(Y) represents the prior of names. • F. Implementation Detail: Face Representation • the appearance of each face is described by local binary pattern (LBP). • the face image is divided into small regions from which the LBP features are extracted and concatenated into a single feature histogram. • pply PCA to reduce the dimension of face descriptor from over 3000 to 500 dimensions. Evaluation