150 likes | 407 Views
Goal : Promote progress in content-based retrieval from digital video ... Collected from Internet Archive and Open Video websites, documentaries from the ...
E N D
Slide 2:Outline Summary of TREC Video Track Automatic Retrieval Task
Our Method and System Architecture
Video Retrieval Demo
Slide 3:2002 TREC Video Retrieval Task Goal : Promote progress in content-based retrieval from digital video via open, metrics-based evaluation.
Query
25 queries
Text, Image(Optional), Video(Optional)
Search Collection
Total Length: 40.16 hours
MPEG-1 format
Collected from Internet Archive and Open Video websites, documentaries from the ‘50s
14,000 shots
292,000 I-frames (images)
Slide 4:Sample Query XML Representation
Slide 5:Sample Query Text : Find pictures of George Washington
Slide 6:System Architecture (Last Year) Simply weighted linear combination of video, audio and text retrieval score
Slide 7:System Architecture (This Year) New step:
Classification through Pseudo Relevance Feedback (PRF)
Combine with movie information agent (abstract, title)
Slide 8:What is Pseudo Relevance Feedback Relevance Feedback (Human intervention)
Slide 9:Classification from Modified PRF Automatic retrieval technique
Modification: use negative data as feedback
Step-by-step
Run base retrieval algorithm on image collection
Nearest neighbor(NN) on color and texture
Build classifier
Negative examples: least relevant images in the collection
Positive examples: image queries
Classify all data in the collection
Slide 10:Combination of Agents Multiple Agents
Text Retrieval Agent
Movie Information Retrieval Agent
Base Image Retrieval Agent
Nearest Neighbor on Color
Nearest Neighbor on Texture
One-class SVM ( not used in TREC )
Classification PRF Agent
Combination of multiple agents
Convert scores to posterior probability
Linear combination
Slide 11:Demo Queries window
Agents (Query Menu)
Slide 12:Demo : Query Expansion through Text Expand the query with google image search engine
Text Queries ? images
Didn’t work as expected (Not used in TREC)
Future work
Slide 13:Demo : Results Menu Show retrieval results (pre-computed)
Answer key window
Results window
Sort by different agents
Rank vs Score
Display results
Slide 14:Demo 5: Result statistics Show result statistics window
Show comparative performance
Ranking in all the participants
Rank 3rd in 27 participating systems
Slide 15:Discussion & Future Work Discussion
The result is sensitive to the queries with small number of answers
Images only is not enough to represent the semantics
Future Work
Incorporate more agents
Utilize the relationship between multiple agent information
Better combination scheme