1. Automatic Video Tagging using Content Redundancy Stefan Siersdorfer 1, Jose San Pedro 2, Mark Sanderson 2 1 L3S Research Center, Germany 2 University of Sheffield, UK SIGIR 2009 2009. 11. 06. Summarized and Presented by Hwang Inbeom, IDS Lab., Seoul National University

2. Copyright  2009 by CEBT Large Amount of Data on YouTube  Traffic to/from YouTube accounts for over 20% of the web total Comprising 60% of on-line watched videos  Growing beyond human perception  Necessity to provide effective knowledge mining and retrieval tools 2

3. Copyright  2009 by CEBT Knowledge Mining and Retrieval  Making use of human annotation: Folksonomy Provides relevant results at a relatively low cost Applications – Topic detection and tracking – Information filtering – Document ranking – Etc.  However, content-based retrieval techniques are not mature enough Folksonomy-based techniques outperform content-based techniques 3

4. Copyright  2009 by CEBT Problem: Poorly Annotated YouTube Videos  Hard to annotate videos Intellectually expensive process Time consuming job  Low-quality tags Often very sparse Lack consistency Present numerous irregularities  Difficult to provide retrieval and knowledge extraction relying on textual features 4

5. Copyright  2009 by CEBT Motivation  Significant amount of near-duplicate videos Over 25% near-duplicate videos detected in search results Has been considered as a problem of online videos  Authors have seen this redundancy as a feature Linkage between two different videos Exploit redundancies to obtain richer video annotations 5

6. Copyright  2009 by CEBT PageRank-like Graph of Videos 6

7. Copyright  2009 by CEBT PageRank-like Graph of Videos 7 Overlap Graph G O = (V O, E O ) Overlap Graph G O = (V O, E O )

8. Copyright  2009 by CEBT Edge in Graph 8  Means video i and j has redundant visual information  Three types of links Duplicate videos Part-of relationship Overlapping Video i Video j

9. Copyright  2009 by CEBT Related Work: VisualRank (WWW 2008)  Builds a graph of images using visual similarity between two images  Runs PageRank algorithm to re- rank images 9

10. Copyright  2009 by CEBT Automatic Tagging  Different approach with that of VisualRank Aims to enrich annotations Not to improve search result  Three methods Simple neighbor-based tagging Overlap redundancy aware tagging TagRank: Context-based tag propagation in video graphs 10

11. Copyright  2009 by CEBT Simple Neighbor-based Tagging  Transforms G O Into the directed graph G’ O (V’ O, E’ O ) of overlapping videos  Weighting function of (i,j) describes to what degree video j is covered by video i 11 Video i Video j w(v i, v j ) w(v j, v i )

12. Copyright  2009 by CEBT Simple Neighbor-based Tagging (contd.)  Gets tag t’s relevance score for a video from information of adjacent videos Weighted sum of influences of overlapping videos tagged by t Counts only adjacent videos’ tags 12 if v j is tagged with t otherwise

13. Copyright  2009 by CEBT An Example 13 t t t t t t t t t’s relevance score

14. Copyright  2009 by CEBT Overlap Redundancy Aware Tagging  Potential high increase of relevance score if a video has multiple redundant overlaps  Contribution of same tag is reduced by relaxation parameter 14

15. Copyright  2009 by CEBT TagRank  Tag weight propagates through the overlap graph  Relevance scores are computed in matrix form TR converges into a certain value: solved with power iteration method Start power iteration with original tagging information and limited number of iteration – To keep original tag relevance – To prevent TR(t) converging into uniform value 15 t t

16. Copyright  2009 by CEBT Evaluation  Two kinds of evaluation: Machine-oriented and human-oriented view Data organization with automatically generated tags – Classification – Clustering User-based evaluation 16

17. Copyright  2009 by CEBT Data Collection  38,283 videos: initial set C Returned videos with top 500 general queries Together with related videos given with results  Redundancy analysis Over 35% of videos (V O ) overlap with one or more other videos 17

18. Copyright  2009 by CEBT Data Organization  Classification with 7 YouTube categories Each of them is containing over 900 videos in V O Binary classification with SVM – Feature vectors constructed with original tags/automatically generated tags Four strategies – BaseOrig: Only considering user-provided tags – NTag: Simple Neighbor-based tagging – RedNTag: Overlap redundancy aware tagging – TagRankΓ: TagRank with Γ iterations 18

19. Copyright  2009 by CEBT Data Organization  Clustering k-Means clustering Partition videos into k categories  Neighbor-based tagging and overlap redundancy aware tagging outperform baseline and TagRank methods in both experiments 19

20. Copyright  2009 by CEBT User-based Evaluation  Assessors rate new tags with web interface Increasingly higher average score when considering tags having higher autotag relevance score 20

21. Copyright  2009 by CEBT Conclusions  Content redundancy in social sharing systems can be used to obtain richer annotations  Additional information obtained by automatic tagging can largely improve automatic organization of content There is information gain for users also  Future work Authors plan to generalize this work to consider different domains – Photos in Flickr – Text in Delicious Analysis and generation of deep tags – Tags linked to a small part of larger media source 21

22. Copyright  2009 by CEBT Discussion  Good idea and good formalization  Would be better if performance of TagRank were good Considering only neighbors is too naïve method  How can we deal with overhead of visual processing?  Would it be scalable enough to apply it to all videos in YouTube? 22