1 Unsupervised Modeling and Recognition of Object Categories with Combination of Visual Contents and Geometric Similarity Links Gunhee Kim Christos Faloutsos Martial Hebert Computer Science Carnegie Mellon University October 31, 2008, Vancouver, Canada ACM MIR 2008
29
Embed
1 Unsupervised Modeling and Recognition of Object Categories with Combination of Visual Contents and Geometric Similarity Links Gunhee Kim Christos Faloutsos.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Unsupervised Modeling and Recognition of Object Categories with Combination of Visual
Contents and Geometric Similarity Links
Unsupervised Modeling and Recognition of Object Categories with Combination of Visual
Contents and Geometric Similarity Links
Gunhee KimChristos Faloutsos
Martial Hebert
Computer ScienceCarnegie Mellon University
October 31, 2008, Vancouver, CanadaACM MIR 2008
2
OutlineOutline
• Problem Statement & Our Approach
• Word Histogram & Network Construction
• pLSA and LDA based Models
• Unsupervised Modeling & Recognition
• Experiments
• Discussion
3
Unsupervised Modeling Unsupervised Modeling
• Category discovery + Ranking
4
Recognition Recognition
Novel Images
Bicycle
Sheep Sign
• Classification + Localization
5
IntuitionIntuition
• Combination of Topic contents and Link AnalysisLatent Topic: Bicycles
Word distributions
Same latent TopicDifferent latent Topic
(Sparse and irregular links)
link distributions
(Dense and consistent links)
link distributions
[1] Sivic, ICCV 2005[2] Fei Fei, ICCV 2005
6
Intuition Intuition
• Combination of Topic contents and link analysis
• Samples of visual words based on Bag-of-Words
• Samples of links generated by image matching
• Two types of evidence into a single generative model
– Ex. Hierarchical Bayesian Models (pLSA, LDA)
7
Our Previous WorkOur Previous Work
• Unsupervised Modeling using Link Analysis Techniques [Kim, CVPR08]
Large Scale Network
Link Analysis Techniques
(ex. PageRank)
- Only links- Only modeling
→ Visual content + Links→ Modeling + Recognition
8
Pros over Conventional Models (1/2)Pros over Conventional Models (1/2)
• Easy Plug-in of geometric information
Indirect Formulation: Link generation with geometric consistency+ Independent of number of parts[Liu, ICCV 2008]