Citation Context

...dict words for a given image by computing the wordssthat have a high posterior probability given the image. Othersmethods attempt to combine large semantic text models withsannotated image structures =-=[7]-=-. Though impressive, suchsalgorithms based on learning don’t work very well in generalssettings and work only marginally well in restricted settings.sFor example, the work described in [7] only gave r...

Citation Context

...es. The most successfulsapproaches learn from large databases of annotated images.sAnnotations typically refer to the contents of the image, andsare fairly specific and comprehensive. Methods such as =-=[2,3]-=-scluster image representations and annotations to produce asjoint distribution linking images and words. These methodsscan predict words for a given image by computing the wordssthat have a high poste...

Citation Context

...uation consisted of searching for images using thesvocabulary words, and only 80 of the words resulted insreasonable images).sA different line of research attempts to find specific objectssin images. =-=[17]-=-, for instance, introduced a method to locateshuman faces in still photographs. These algorithms arestypically accurate, but have not been developed for a widesrange of objects. Additionally, combinin...

Citation Context

...es. The most successfulsapproaches learn from large databases of annotated images.sAnnotations typically refer to the contents of the image, andsare fairly specific and comprehensive. Methods such as =-=[2,3]-=-scluster image representations and annotations to produce asjoint distribution linking images and words. These methodsscan predict words for a given image by computing the wordssthat have a high poste...

Citation Context

...Web. The labels produced using our game cansbe combined with these techniques to provide a powerfulssolution.sThe Open Mind InitiativesOur work is similar in spirit to the Open Mind Initiative (e.g.,s=-=[18,19]-=-), a worldwide effort to develop “intelligent”ssoftware. Open Mind collects information from regularsInternet users (referred to as “netizens”) and feeds it tosmachine learning algorithms. Volunteers ...

Citation Context

...the contents ofsimages on the Web are related to the text appearing in thespage. This is insufficient because the text adjacent to thesimages is often scarce, and can be misleading or hard tosprocess =-=[4]-=-.sThe only method currently available for obtaining precisesimage descriptions is manual labeling, which is tedious andsthus extremely costly. But, what if people labeled imagesswithout realizing they...

Citation Context

...e re-inserted into the game whensseveral months have passed because the meaning of thesimages may have changed due to maturation effects. ThesEnglish language changes over time, as do other languagess=-=[20]-=-. We want to capture the labels appropriate to an image,sand thus if the language referring to that image changes overstime, so should our labels. In addition to changes inslanguage, cultural changes ...

Citation Context

...Web. The labels produced using our game cansbe combined with these techniques to provide a powerfulssolution.sThe Open Mind InitiativesOur work is similar in spirit to the Open Mind Initiative (e.g.,s=-=[18,19]-=-), a worldwide effort to develop “intelligent”ssoftware. Open Mind collects information from regularsInternet users (referred to as “netizens”) and feeds it tosmachine learning algorithms. Volunteers ...

Citation Context

...is probably meaningful. In fact,ssince these labels come from different people, they have thespotential of being more robust and descriptive than labelssthat an individual indexer would have assigned =-=[15]-=-.sTo increase the probability that a label for a particular imagesis meaningful, we utilize a good label threshold. This meanssPlayer 1 guesses: pursesPlayer 1 guesses: bagsPlayer 1 guesses: brownsSuc...