Date post: | 20-Jan-2017 |
Category: |
Technology |
Upload: | benoit-huet |
View: | 813 times |
Download: | 1 times |
Maria Eskevich, Quoc-Minh Bui, Hoang-An Le, Benoit Huet
Multimedia DepartmentEURECOM
Sophia-Antipolis, France
Exploring Video Hyperlinking in
Broadcast Media
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 2
Motivation CURRENTLY: Constantly growing quantity of
multimedia content produced by both professionals and individual users.
NEED: navigation systems that allow access to this data on different levels of granularity in order to contribute to further discovery of a topic of interest for the user or to facilitate individual user browsing within a collection.
10/30/15
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 3
Task and Challenges We envisage the users to be interested in creating
their own path within the multimedia collection based on their level of awareness/knowledge of it, and tasks.
How do we create a hyperlinked collection to allow each individual user to follow their own path of interest within?
Why is it challenging? Lack of interpretability of the content: Overall textual description on the video level which leads to:
Retrieval process is based on text features, while the rich multimodality of videos is not exploited;
Not possible to partially retrieve a video, a specific fragment without having to watch the entire video.
Archives are not static: Need for dynamic creation of links
10/30/15
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 4
Insights in Hyperlinking Hyperlinking
Creating “links” between media
Video Hyperlinking video to video video fragment to video fragment
10/30/15
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 5
Related experiments: Search and Hyperlinking @ MediaEval/TRECVid Different techniques based on: Video segmentation method:
Fixed-length units (same length, sentences, speech segments) with further adjustment to match full sentences, using speech segment boundaries and pauses. they build a probabilistic framework to model the importance of words and refine segment boundaries accordingly;
Meaningful segments, based on topics derived from transcripts: classification trees to define the starting and ending times of these segments; lexical cohesion within segment.
Features used for retrieval: Text similarity (vector-based models and TF-IDF weightings); Named entities or synonyms; Visual information: visual concepts or SURF and SIFT features,
detected at the shot level.
10/30/15
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 6
Search and Hyperlinking @ MediaEval/TRECVid Our solution: Scene segmentation technique based on
visual and temporal coherence of the video segments that constitute the scenes of the video.
Experiment with hybrid segmentationVisual, Topic and Temporal Coherence
Use visual features to improve the ranking of results of the hyperlinking including visual analysis during the search process.
10/30/15
System overview
Broadcast Media: BBC contentManually transcribed
subtitlesMetadata:
title, cast, description, broadcast time
Shot segmentation, keyframes
Visual content analysis:
Scene segmentation
Concept detection
on shot level
Lucene/Solr:Indexing/Retrieval on shot
level
Media database:76 214 fragments
Webservice:(HTML5/AJAX/PHP)
Videos:2323 items/1697 hours
User Interface
Shot segmentation, keyframe extraction
Result list:Media fragments
level
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 8
151 Visual Concepts (TrecVid 2012) 3_Or_More_People Actor Adult Adult_Female_Human Adult_Male_Human Airplane Airplane_Flying Airport_Or_Airfield Anchorperson Animal Animation_Cartoon Armed_Person Athlete Baby Baseball Basketball Beach Bicycles Bicycling Birds Boat_Ship Boy
• Building• Bus• Car• Car_Racing• Cats• Cattle• Chair• Charts• Child• Church• City• Cityscape• Classroom• Clouds• Construction_Vehicles• Court• Crowd• Dancing• Daytime_Outdoor• Demonstration_Or_Protest• Desert• Dogs
• Emergency_Vehicles• Explosion_Fire• Face• Factory• Female-Human-Face-Closeup• Female_Anchor• Female_Human_Face• Female_Person• Female_Reporter• Fields• Flags• Flowers• Football• Forest• Girl• Golf• Graphic• Greeting• Ground_Combat• Gun• Handshaking• Harbors
• Helicopter_Hovering• Helicopters• Highway• Hill• Hockey• Horse• Hospital• Human_Young_Adult• Indoor• Insect• Kitchen• Laboratory• Landscape• Machine_Guns• Male-Human-Face-Closeup• Male_Anchor• Male_Human_Face• Male_Person• Male_Reporter• Man_Wearing_A_Suit• Maps• Meeting• …
10/30/15
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 9
Scene segmentation Scene : group of shots based on content
similarity and temporal consistency among shots; Content similarity : visual similarity between HSV
histograms extracted from the keyframes of different shots;
Grouping is performed using two extensions of the Scene Transition Graph (STG): reduces the computational cost of STG-based shot grouping
by considering shot linking transitivity, builds on the former to construct a probabilistic framework
that alleviates the need for manual STG parameter selection. [Apostolidis et al. “Automatic fine-grained hyperlinking of videos within a
closed collection using scene segmentation.” ACM MM 2014]
10/30/15
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 10
Handling Visual Concepts in Solr
10/30/15
http://localhost:8983/solr/collection_mediaEval/select?q=text:(Children out on poetry trip Exploration of poetry by school children Poem writing) Animal:[0.2 TO 1] Building:[0.2 TO 1]
Schema = structure of document using fields of different types
Query
<doc><field name="id">
20080401_013000_bbcfour_legends_marty_feldman_six_degrees_of#t=399,402</field><field name="begin">00:06:39.644</field><field name="end">00:06:42.285</field><field name="videoId">20080401_013000_bbcfour_legends_marty_feldman_six_degrees_of</field><field name="subtitle">'It was very, very successful.'</field>
<field name="Actor">0.143</field><field name="Adult">0.239</field><field name="Animal">0.0572</field>
</doc>
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 11
Results (S&H 2014)
10/30/15
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 12
Results discussion Text (Audio) performs best on the 2014
MediaEval Hyperlinking task Using visual features isn’t
straightforward
The ‘MoreLikeThis’ approach is outperformed by the Text only searchPossibly due to query formulation differences
Sentences vs Keywords Visual Scenes outperform other
fragmentation levels (Topic, Sentences)10/30/15
[Safadi et al. “When textual and visual information join forces for multimedia retrieval”, ICMR 2014]
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 13
DEMO Hyper Video Browser
10/30/15
Search/Browse Hyperlinking
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 14
Conclusion and Future work We proposed and evaluated an approach to include
visual properties in the search of video segments for hyperlinking.
Experimental results show that : Visual properties provide meaningful cues for segmenting the video Mapping text-based queries to visual concepts is not easy Automatic selection of relevant concepts is required (human
intervention is impractical and does not necessarily lead to perfect results)
Operating at the keyword level rather than full text offers improvements
Current/Future work: incorporate query semantics when identifying key visual semantic concepts based on named entity recognition approaches and keyword/visual
concepts co-occurrences
10/30/15
SLAM Workshop @ ACM MM 2015, Brisbane, Australia 15
References E. Apostolidis, V. Mezaris, M. Sahuguet, B. Huet, B. Cervenkova, D. Stein, S.
Eickeler, J. L. Redondo Garcia,R. Troncy, and L. Pikora. “Automatic ne-grained hyperlinking of videos within a closed collection using scene segmentation”. In ACMMM 2014, 22nd ACM International Conference on Multimedia, Orlando, Florida, USA, 11 2014.
H. Le, Q. Bui, B. Huet, B. Cervenkova, J. Bouchner, E. Apostolidis, F. Markatopoulou, A. Pournaras, V. Mezaris, D. Stein, S. Eickeler, and M. Stadtschnitzer,.“LinkedTV at MediaEval 2014 Search and Hyperlinking Task”. In Proceedings of MediaEval 2014 Workshop, 2014.
R. J. F. Ordelman, M. Eskevich, R. Aly, B. Huet, and G. J. F. Jones. “Dening and evaluating video hyperlinking for navigating multimedia archives”. In Proceedings of the 24th International Conference on World Wide Web Companion, WWW 2015, Florence, Italy, May 18-22, 2015 - Companion Volume, pages 727-732, 2015.
M. Sahuguet, B. Huet, B. Cervenkova, E. Apostolidis, V. Mezaris, D. Stein, S. Eickeler, J. L. Redondo Garcia, and L. Pikora. “LinkedTV at MediaEval 2013 search and hyperlinking task”. In MediaEval 2013 Workshop, Barcelona, Spain, 10 2013.
B. Safadi, M. Sahuguet, and B. Huet. “When textual and visual information join forces for multimedia retrieval”. In Proceedings of International Conference on Multimedia Retrieval (ICMR '14). ACM, New York, NY, USA, , Pages 265 , 8 pages, 2014.
10/30/15