Software Engineering Institute April 29, 2009

1. Informedia Digital Video Library Research Over 15 Years: Enabling Better Access to Voluminous Multimedia Interfaces Mike Christel christel@cs.cmu.edu Entertainment Technology CenterCarnegie Mellon University Software Engineering Institute April 29, 2009

2. Talk Outline • Informedia origins (my journey from SEI to Wean Hall) • Creating metadata for video information sets • Informedia demonstrations • oral history collection • news video collection • Types of search: beyond fact-finding • Exploratory search through multiple views • Evaluation hurdles • Reflections on Informedia work

3. Informedia Origins (from SEI to Wean Hall) • CMU/SEI-93-TR-012, AMORE: The Advanced Multimedia Organizer for Requirements Elicitation • CMU Pres. Mehrabian, Media Arts and Technology Alliance with WQED Pittsburgh • Speech recognition (Raj Reddy) • Computer vision/image processing (Takeo Kanade) • Language technologies (Jaime Carbonell) • Founding of CMU Human-Computer Interaction Institute • NSF Digital Libraries Initiative Phase 1 (1994)

4. CMU Informedia Digital Video Research • Details at: http://www.informedia.cs.cmu.edu • Speech recognition and alignment • Image processing • Named entity tagging • Synchronized metadata for search and navigation • Fast, direct video access to oral histories, news, etc. • Demonstration oral history corpus: 913 hours of interviews from 400 individuals, 18,254 interview story segments (average story segment length of 3 minutes) • Demonstration news corpus: TRECVID 2006 test set (165 hours of U.S., Arabic, and Chinese news with 79,484 reference shots)

5. The HistoryMakers Oral History Archive • http://www.thehistorymakers.com • Funded in part by IMLS Grant LG-03-03-0048-03 • The world’s largest African American oral history archive with accomplished African Americans • Purpose: • To educate and show the breadth and depth of this important American history as told by the first person • To highlight the accomplishments of individual African Americans across a variety of disciplines • To preserve this material for generations to come • Committed to exposing the archive to the widest audience possible, making use of new technologies as appropriate

6. A Theme for Today: User Involvement • User Correction: Corrective action for metadata errors (analogous to Harry Shum’s vision at Microsoft for human-assisted computer vision success) • User Control: Driving the interface to overcome metadata errors • User Context: More useful interfaces driven implicitly by context

7. Speech Recognition Functions • Generates transcript (if one is not given) to enable text-based retrieval from spoken language documents • Improves text synchronization to audio/video in presence of scripts (align speech with text) • Supplies necessary information for library segmentation and multimedia abstractions (e.g., break stories apart at silence points rather than in the middle of sentences)

8. Speech Alignment Example

9. Image Understanding Functions • Scene segmentation • Similarity matching • Camera motion determination and object tracking • Optical Character Recognition (OCR) on video text and titles • Face detection and recognition • Ongoing research work in object identification and scene characterization, e.g., indoor/outdoor, road, building, etc.

10. Images Containing Similar Colors… Image search with tropical rainforest image leads to…

11. Images Containing Similar Colors

12. Images Containing Similar Shapes

13. Images Containing Similar Content

14. Static Static Zoom Adult Female Animal Two adults Head Motion Left Motion None CNN LIVE CNN An Online First Studio Outdoor Indoor Goal: Automatic Video Characterization Shots Yellowstone Camera Objects Action Captions Scenery

15. Static Static Zoom Adult Female Animal Two adults Head Motion Left Motion None CNN LIVE CNN An Online First Studio Outdoor Indoor Goal: Automatic Video Characterization Shots Yellowstone Camera Objects Action Captions Scenery

16. Automated Video Processing • Produces descriptive metadata for video libraries • Metadata has errors greater than metadata produced by a careful, human-provided annotation • Errors in metadata can be reduced – examples to follow…

17. Camera and Motion Detection Pan Right object motion (not pan left)

18. Text and Face Detection

19. Video OCR Block Diagram Text Area Detection Video Text Area Preprocessing Commercial OCR ASCII Text

20. Video FramesFiltered Frames AND-ed Frames (1/2 s intervals)

21. VOCR Preprocessing Problems

22. Augmenting VOCR with Dictionary Look-up

23. Named Entity Extraction F. Kubala, R. Schwartz, R. Stone, and R. Weischedel, “Named Entity Extraction from Speech”, Proc. DARPA Workshop on Broadcast News Understanding Systems, Lansdowne, VA, February 1998. CNN national correspondent John Holliman is at Hartsfield International Airport in Atlanta. Good morning, John. …But there was one situation here at Hartsfield where one airplane flying from Atlanta to Newark, New Jerseyyesterday had a mechanical problem and it caused a backup that spread throughout the whole system because even though there were a lot of planes flying to the New York area from the Atlanta area yesterday, …. Key: Place, Time, Organization/Person

24. MetadataExtractor People Event Affiliation Location Time Topics User Interface Perspective Templates (final representation) Summarizer Enhancing Library Utility via Better Metadata

25. Improving the Interface via Usage Context Example: query-based thumbnail selection

26. Improving Utility through End-User Control Example: filtering storyboard based on visual concepts with user controlling precision and recall

27. Improving the Metadata via User Interaction • Example: collecting positive and implicit negative sets of labeled shot data for visual concepts • Reference: Ming-yu Chen, et al., ACM Multimedia 2005

28. Automated Video Processing Produces descriptive metadata for video libraries Metadata has errors greater than metadata produced by a careful, human-provided annotation Errors in metadata can be reduced: By more computation-intensive algorithms By taking advantage of video frame-to-frame redundancy By folding in context, e.g., probable text sizes in video By folding in extra sources of knowledge, e.g., a dictionary for cleaning up VOCR, or labeled data revealing patterns for named entity detection By human review and correction, which can generate additional labeled data for machine learning

29. User Involvement • User Correction: Corrective action for metadata errors (analogous to Harry Shum’s vision at Microsoft for human-assisted computer vision success) • User Control: Driving the interface to overcome metadata errors • User Context: More useful interfaces driven implicitly by context

30. Video Summaries (without User Context) • BBC rushes video summarization task in TRECVID 2007 and TRECVID 2008 shows difficulty of the task • Video summary is “a condensed version of some information, such that various judgments about the full information can be made using only the summary and taking less time and effort than would be required using the full information source” • Maximum 4% duration (2% in TRECVID 2008) • Benefits of this TRECVID task: provides a reasonably large video collection to be summarized, a uniform method of creating ground truth, and a uniform scoring mechanism

31. BBC Rushes • 42 test videos (+ development ones) from BBC Archive • Test videos: • minimum duration 3.3 minutes • maximum duration 36.4 minutes • mean duration 25 minutes • Raw (unedited) rush video with a great deal of redundancy (repeated takes), mixed quality audio, “junk” frames

32. Video Summaries (with/without User Context) • BBC Rush video has no context to build from • However, users often provide cues as to what is important, as will be seen shortly

33. Storyboards: TRECVID Search Success • For the shot-based directed search information retrieval task evaluated at TRECVID, storyboards have consistently and overwhelmingly produced the best performance • Motivated users can navigate through thousands of shot thumbnails in storyboards, better even than with “extreme video retrieval” interfaces: 2487 shots on average per 15 minute topic for TRECVID 2006 (Christel & Yan, CIVR 2007) • Storyboard benefits: packed visual overview, trivial interactive control needed for “overview, zoom and filter, details on demand” – Shneiderman’s Visual Information-Seeking Mantra

34. Beyond Fact-Finding • CACM April 2006 special issue on this topic • G. Marchionini (“Exploratory Search: From Finding to Understanding,” CACM 49, April 2006) breaks down 3 types of search activities: • Lookup (fact-finding; solving stated/understood need) • Learn • Investigate • Computer scientists and information retrieval specialists emphasize evaluation of lookup activities (NIST TREC) • Real world interest in learn/investigate: for an oral history collection, State Univ. New York at Buffalo Workshop library science and humanities participants quite interested in learn/investigate activities

35. Exploratory Search • Examples where storyboards still useful: visual review, e.g., of disaster field footage • Where storyboards fail: • Showing other facets like time, space, co-occurrence, named entities (When did disasters occur? Which ones? Where?) • Providing collection understanding, a holistic view of what’s in say 100s of segments of 1000s of matching shots • Providing window into visually homogenous results, e.g., results from color search perhaps, or a corpus of just lecture slides, or head-and-shoulder interview shots • Claim: Storyboards are not sufficient, but are part of a useful suite of tools/interfaces for interactive video search

36. Anecdotal Support for Claim • Collected 2006-2007 from: • Government analysts with news data • History students and faculty with oral history data • Views Tested: • Timeline • Visualization By Example (VIBE) Plot (query terms) • Map View • Named Entity view (people, places, organizations) • Text-dominant views: • Nested Lists (pre-defined clusters by contributor) • Common Text (on-the-fly grouping of common phrases)

37. Anecdotal Results • 38 HistoryMakers corpus users (mostly students, 15 female, average age 24), experienced web searchers, modest digital video experience • 6 intelligence analysts (1 female; 2 older than 40, 3 in their 30s, 1 in 20s), very experienced text searchers, experienced web searchers, novice video searchers • View use minimal aside from Common Text • Text titling and text transcripts used frequently • A bit of evidence for collection understanding (e.g., diffs in topic between New York and Chicago), but overall, cautious use of default settings for initial trial(s).

38. Evaluation Hurdles • How does one evaluate information visualization for promoting exploratory video search? • Low level simple tasks vs. complex real-world tasks • Traditional effectiveness, efficiency, satisfaction are even problematic: is “fast” interface for exploration good or bad? • HCI discount usability techniques offer some support, but ecological validity may limit impact of conclusions (e.g., HCII students found Common Text well suited for History students) • Look to field of Visual Analytics for help, e.g., Plaisant • “First hour with system” studies, or “developer as user” insights too limiting. Rather, consider Multi-dimensional In-depth Long-term Case-studies (MILC)

39. Reflections – Some Informedia Successes • Open benchmarking to gauge progress in digital video libraries and video information retrieval • NIST TREC Spoken Document Retrieval • NIST TRECVID • Application of machine learning techniques for visual classification; addressing “semantic gap” through visual concepts (Rong Yan, Wei-Hao Lin, Jun Yang, Robert Chen with Alex Hauptmann as PhD thesis advisor in LTI) • User studies to empirically drive interface development (see http://www.morganclaypool.com/toc/icr/1/1 - Morgan & Claypool Synthesis Lectures on Information Concepts, Retrieval, and Services)

40. Reflections – Opportunities Missed • Foreseeing growth of the Web in 1994 • Limited use and dissemination via work with broadcasters and intellectual property concerns • Significant shifts in environment, e.g., from $1,000,000 for a terabyte of storage in 1994 to $100 (or less) today • Emphasis on information retrieval in traditional sense (lookup tasks)

41. Conclusions to Build Upon - 1 “Interactive” allows human direction to compensate for automation shortcomings and varying needs Interactive fact-finding better than automated fact-finding in visual shot retrieval (TRECVID) Interactive computer vision has successes (Harry Shum at Microsoft, Michael Brown et al. at NUS) Interactive view/facet control == ??? (too early to tell) Users need scaffolding/support to get started Evaluations need to run longer term, in depth, with case studies to see what has benefit (MILC)

42. Conclusions to Build Upon - 2 Storyboards work well for visual overview Video surrogates can be made more effective, efficient, and satisfying when tailored to user activity (leverage context) Interface should provide easy tuning of precision vs. recall As cheap storage and transmission is producing a wealth of digital video, exploratory search will gain emphasis regarding video repositories Augment automatically produced metadata with human-provided descriptors (take advantage of what users are willing to volunteer, and in fact solicit additional feedback from humans through motivating games that allow for human computation, a research focus of Luis von Ahn at Carnegie Mellon University)

43. Informedia Future Work • Video surveillance – Informedia has processed a month of data from 24 cameras in nursing home (45 terabytes) • Leveraging from the Web (YouTube mining) • Exploring utility of video for engagement and entertainment with education focus (Entertainment Technology Center - ETC) • Opening up access to tools and interface (web deployment; open-sourcing “Informedia Core”; opportunities exist – Oral History Association Meeting) • Synthetic Interviews (with Scott Stevens at the ETC)

44. Credits Many members of the Informedia Project and CMU research community contributed to this work, including: Informedia Project Director: Howard Wactlar Informedia Library Technical Essentials: Bob Baron, Bryan Maher Informedia Administration: Colleen Everett, Melissa Keaton, Scott Stevens Informedia User Interface: Ron Conescu, Neema Moraveji, Chang Huang Informedia Processing: Alex Hauptmann, Ming-yu Chen, Wei-Hao Lin, Rong Yan, Jun Yang, Rong Jin, Michael Smith, Michael Witbrock HistoryMakers Beta Testers: Joe Trotter (CMU History Dept.), SUNY at Buffalo and all UB Workshop participants: Schomburg Center for Research in Black Culture, NY Public Library, Randforce Associates, University of Illinois (3 campuses), Drexel University • This work supported by the National Science Foundation under Grant Nos. IIS-0205219, IIS-0535056, and IIS-0705491.