Transcript Yung-Yu Chuang

Research 2.0
Harnessing Collective Intelligence
Yung-Yu Chuang 莊永裕
Communication & Multimedia
Laboratory
National Taiwan University
Research 2.0
• Research 2.0 = Research based on the
concept of Web 2.0
• Similar idea/term was proposed by Harry
Shum of MSRA
• Observations from vision and multimedia
research
Web 2.0
Web2.0的精神在於”肯定網路上不特定多數人並非
被動的服務享受者，而是主動的創作者，並積極地
開發技術或服務，鼓勵這些人參與。”
梅田望夫
Web 1.0
Web 2.0
DoubleClick
Google AdSense
mp3.com
Napster
Britannica online
wikipedia
personal website
blogging
publishing
participation
The long tail
80-20 rule
Law of the vital few
Books, media, software…
Web 2.0 involves all people
and shifts the authority.
Web 2.0 (Tim O’Reilly)
•
•
•
•
The web as platform
Data is the next Intel Inside
Harnessing collective intelligence
…
Research 2.0
• Data, paper and code are on the web
– Benchmark becomes more and more
important. Sharing your data and code is
likely to make your research more influential.
Stereo problem
Middlebury stereo page
Middlebury stereo page
Performance for over 40 methods were reported;
36 of them were submitted by other researchers.
Middlebury stereo page
• A review paper along with a benchmark
was published in IJCV 2002.
• 541 citations since then according to
Google scholar.
LIBSVM (C.J.Lin at NTU)
• 873 citations since 2001 according to
Google scholar.
• SVM is not necessarily the best tool for
classification.
• Its popularity could gain from some
robust and easy-to-use tools.
Research 2.0
• Data, paper and code are on the web
– Benchmark becomes more and more
important. Sharing your data and code is
likely to make your research more influential.
Research 2.0
• Data, paper and code are on the web
– Benchmark becomes more and more
important. Sharing your data and code is
likely to make your research more influential.
• Explore vast amount of (noisy) data
– Statistical approaches (machine learning,
data mining, information retrieval)
Landmark project
• What are the text keywords for landmarks?
• What are the visual keywords associated with
landmarks?
Research 2.0
• Data, paper and code are on the web
– Benchmark becomes more and more
important. Sharing your data and code is
likely to make your research more influential.
• Explore vast amount of (noisy) data
– Statistical approaches (machine learning,
data mining, information retrieval)
• Utilize collective intelligence
– Good designs and motivations encourage
people to make contributions
What can users contributes?
•
•
•
•
•
•
YouTube/flickr: media and tags
Wikipedia: knowledge
Amazon: reviews/comments
Connextions: courses
MIT’s openmid: common sense
Human computation cycles
Application to ROI
•
•
•
•
We have applied this idea to ROI research.
There is no benchmark
There is no evaluation
There is no example-based approach
What is ROI?
How to detect?
• Heuristics
– Contrast
– Face
– Text
– Shape
…
How to detect?
• Heuristics
– Contrast
– Face
– Text
– Shape
…
• User labeling
– Manual
– Eye tracker
…
Our approach
• Collect large amount of ground truth
• Evaluate existing algorithms
• A learning-based algorithm
Conclusions
Because of Internet’s paradigm shift, what
are new research possibilities? The answers
are left to you.