Data mining for Web Video Auto Tagging TRAN Hoang Tung.

8
Data mining for Data mining for Web Video Auto Web Video Auto Tagging Tagging TRAN Hoang Tung

Transcript of Data mining for Web Video Auto Tagging TRAN Hoang Tung.

Page 1: Data mining for Web Video Auto Tagging TRAN Hoang Tung.

Data mining for Data mining for Web Video Auto Web Video Auto

TaggingTaggingTRAN Hoang Tung

Page 2: Data mining for Web Video Auto Tagging TRAN Hoang Tung.

OutlineOutlineData mining for Web Video Auto Data mining for Web Video Auto

TaggingTaggingIntroducing myselfIntroducing myself

General ContextGeneral ContextExplain what I want to do in my PhD Explain what I want to do in my PhD thesisthesis

Data miningData mining

Applying data mining techniques into Applying data mining techniques into video taggingvideo tagging

Last year and future plan!Last year and future plan!

USTH Consortium - Toulouse 2011USTH Consortium - Toulouse 2011 10/19/1110/19/1122

Page 3: Data mining for Web Video Auto Tagging TRAN Hoang Tung.

About myselfAbout myselfMy name is TRAN Hoang TungMy name is TRAN Hoang Tung

Arrival date: 25Arrival date: 25thth Oct 2010 Oct 2010

I’m working in Hubert Curien Laboratory, Jean I’m working in Hubert Curien Laboratory, Jean Monnet University and CNRSMonnet University and CNRS

City: Saint Etienne (close to Lyon and City: Saint Etienne (close to Lyon and Grenoble)Grenoble)

My supervisors: My supervisors: Francois Jacquenet (full professor)Francois Jacquenet (full professor)Elisa Fromont (assistant professor)Elisa Fromont (assistant professor)Baptiste Jeudy (assistant professor)Baptiste Jeudy (assistant professor)

Keywords: data mining, video tagging, video Keywords: data mining, video tagging, video annotationannotation

USTH Consortium - Toulouse 2011USTH Consortium - Toulouse 2011 10/19/1110/19/1133

Page 4: Data mining for Web Video Auto Tagging TRAN Hoang Tung.

USTH Consortium - Toulouse 2011USTH Consortium - Toulouse 2011 10/19/1110/19/1144

Page 5: Data mining for Web Video Auto Tagging TRAN Hoang Tung.

Web videos Web videos (Youtube…)(Youtube…)

Current video search engines are Current video search engines are text-text-based based (title, description, tags). Title & (title, description, tags). Title & description are written by each uploader description are written by each uploader (normally as a complete phase). Tags are (normally as a complete phase). Tags are single wordssingle words!!!!

However, tags are notoriously:However, tags are notoriously:Incomplete (don’t fully represent such video)Incomplete (don’t fully represent such video)

Incorrect (spam, increase number of view)Incorrect (spam, increase number of view)

Unranked (the most important tag is not the Unranked (the most important tag is not the first tag)first tag)

My thesis goal: creating an auto-tagging My thesis goal: creating an auto-tagging system which reduces above disadvantages system which reduces above disadvantages of current tags.of current tags.USTH Consortium - Toulouse 2011USTH Consortium - Toulouse 2011 10/19/1110/19/1155

Page 6: Data mining for Web Video Auto Tagging TRAN Hoang Tung.

Data miningData miningData mining is a field of computer Data mining is a field of computer science, and more precisely of science, and more precisely of artificial artificial intelligenceintelligence. The goal is to describe . The goal is to describe (very) large data in an informative way. (very) large data in an informative way. For example For example discovering patterns discovering patterns !!

Example: Market-Basket AnalysisExample: Market-Basket Analysis

USTH Consortium - Toulouse 2011USTH Consortium - Toulouse 2011 10/19/1110/19/1166

TID

items

1 {bread, milk}

2 {bread, diaper, beers, eggs}

3 {bread, diaper, beers, cola}

4 {bread, diaper, beers, milk}

5 {bread, diaper, cola, milk}

Consider item set {bread, diaper}:

Support = 4/5 = 80%

Consider association rule:

{bread, diaper} -> beers

with confidence:

= s({bread,diaper,beers})/s({bread,diaper}) =3/4

Page 7: Data mining for Web Video Auto Tagging TRAN Hoang Tung.

Data mining & video Data mining & video taggingtagging

Assumption: similar videos will have (with Assumption: similar videos will have (with high probability) similar tags.high probability) similar tags.

Steps:Steps:Compute similarities between videos based on Compute similarities between videos based on patternspatterns

Propagate the tagsPropagate the tags

USTH Consortium - Toulouse 2011USTH Consortium - Toulouse 2011 10/19/1110/19/1177

Page 8: Data mining for Web Video Auto Tagging TRAN Hoang Tung.

The Past and The Past and Future!Future!Last year:Last year:

Studying data mining Studying data mining Followed a master course in Machine Followed a master course in Machine LearningLearningParticipating to a winter school in Machine Participating to a winter school in Machine Learning applied to image processingLearning applied to image processing

Reading bibliography about data mining Reading bibliography about data mining applied to video analysisapplied to video analysisStudying French ! (240 hours)Studying French ! (240 hours)

Next year:Next year:Take a look at other types of Take a look at other types of patternspatterns: : discriminative patterns, sequences, …discriminative patterns, sequences, …Try other techniques to convert video into Try other techniques to convert video into binary formbinary formWrite something ! Write something !

USTH Consortium - Toulouse 2011USTH Consortium - Toulouse 2011 10/19/1110/19/1188