ARCOMEM Training Material Understanding Images and Video
Jan 15, 2015
ARCOMEM Training Material
Understanding Images and Video
ARCOMEM Training Material
Introduction
• The web is increasingly multimedia in nature• We concentrate on the multimedia aspects• Image and video content analysis, indexing,
matching and annotation• Exploit the web, social web and linked data
web
ARCOMEM Training Material
Exploring image and video reuse on the Web
ARCOMEM Training Material
• Goal– Use image analysis techniques to aggregate social contexts; in
particular to interlink the social discussion of events and topics through the content of images and video.
• Motivation– Media is often reused or reposted on social networks.
– Detection of near-duplicate multimedia artifacts provides a means to investigate and explore many facets about the documents the media is embedded within. • Some examples include:
– Aggregating documents about the same subject/event/opinion» Finding cases where media is used in differing contexts is also interesting.
– Exploring how different social groups talk about the same media
ARCOMEM Training Material
ARCOMEM Training Material
ARCOMEM US-Elections Crawl Examples
ARCOMEM Training Material
Images
• image/gif 4371• image/jpeg 667073• image/png 46644• image/svg+xml 1• image/x-icon 2• image/x-ms-bmp 167
ARCOMEM Training Material
122 Detected Duplicates
https://sphotos-a.xx.fbcdn.net/hphotos-ash3/63089_10151299959267177_1113898282_n.jpghttp://p.twimg.com/A7Oq4JRCEAA05uG.jpghttps://p.twimg.com/A7RMKgrCIAAvPdj.jpghttp://24.media.tumblr.com/tumblr_md4ttulK8d1qa5ex8o1_500.jpghttps://p.twimg.com/A7Oq4JRCEAA05uG.jpghttps://p.twimg.com/A7JYdoDCEAAktAb.jpghttps://p.twimg.com/A7JIdp7CcAAjsPF.jpghttp://p.twimg.com/A7JSUYCCUAEzW5O.jpghttp://p.twimg.com/A7NoTNOCIAEQ3ie.jpghttps://p.twimg.com/A7LsnVrCQAAWYFA.jpg...
112 Detected Duplicates
http://25.media.tumblr.com/tumblr_mc0mmx2kTA1rxxq3ro1_500.pnghttp://p.twimg.com/A7Ep0CjCMAAon-f.jpghttp://p.twimg.com/A7N3lxrCAAE_gzP.jpghttps://p.twimg.com/A5ctYtVCMAE8_8e.jpghttp://p.twimg.com/A7NMvVACUAEL9aa.jpghttp://p.twimg.com/A7EkZK8CEAE0pYu.jpghttps://p.twimg.com/A7FOSwtCUAADws1.jpghttp://p.twimg.com/A7EiFd4CcAAZdT1.jpghttp://p.twimg.com/A7JwHHbCQAAvqNN.jpg...
ARCOMEM Training Material
103 Detected Duplicates
https://p.twimg.com/A7K8WWiCAAE9Uap.pnghttps://p.twimg.com/A7Iq1DSCQAAyRBG.jpghttp://p.twimg.com/A7L7bQJCIAIaaqm.jpghttp://p.twimg.com/A7KSC5gCcAAcLTj.jpghttps://p.twimg.com/A7OqzarCYAA8L9N.jpghttps://p.twimg.com/A7OlKn-CQAA-j7X.jpghttps://p.twimg.com/A7J5CHPCUAEp-86.pnghttps://p.twimg.com/A7JOd9kCcAIy3YM.jpghttp://25.media.tumblr.com/tumblr_lyr1v4kp8J1qcjsjlo1_500.jpg...
106 Detected Duplicates
https://p.twimg.com/A7MQiKeCMAE1rmX.jpghttps://p.twimg.com/A7IFNt_CUAAqYFV.jpghttps://p.twimg.com/A7DrPLMCIAAoKb8.jpghttp://p.twimg.com/A7Y5BW9CQAAqF3E.jpghttps://p.twimg.com/A2YzbudCQAE77pr.jpghttps://p.twimg.com/A7FQF7PCUAAgWWH.jpghttps://p.twimg.com/A7EO5YFCMAAMTxB.jpghttp://p.twimg.com/A7IWHTtCUAIAtH8.jpghttp://p.twimg.com/A7FcyZdCUAAO_Vg.jpghttp://25.media.tumblr.com/tumblr_m7c9nivzwF1qfep67o1_500.jpg...
ARCOMEM Training Material
ARCOMEM Training Material
Multimedia Opinion Mining
ARCOMEM Training Material
Goals
• Investigate the use of facial analysis to classify facial expressions in images and videos found on the web.– Can be used to indicate emotion of subject.
• Investigate course-grained automatic classification using image features– For abstract opinion-related concepts
(sentiment/privacy/attractiveness)
• Investigate correlations between images and opinions mined from text– Does the same image get reused in different
documents to illustrate the same (or different) opinion?
ARCOMEM Training Material
Sentiment/privacy/attractiveness
• Experimental support for classifying visual media with respect to sentiment, privacy and attractiveness is being built into ARCOMEM.
ARCOMEM Training Material
Image-opinion correlation
• Correlations between images and opinions extracted from the text can be explored by querying the ARCOMEM database.
+ve -ve
ARCOMEM Training Material
Towards Facial Analysis in the Wild
• Detection and analysis of faces in multimedia content can help us guide and contextualise a crawl:– Recognition and expression analysis can help us
determine if an image is relevant or interesting.– Post-crawl the information can be used for visualisation.
• Current research very much based on images taken in lab-conditions; how far can we take it?
ARCOMEM Training Material
Applying the model to crawled images