Webly-Supervised Visual Concept Learning


Webly-Supervised Visual Concept Learning
Learning Everything About Anything
Webly-Supervised Visual Concept Learning
Santosh Divvala
Ali Farhadi
How can we learn everything about anything?
Carlos Guestrin
Results: Relationships discovered
Key Challenges:
• How to gather training data
(queries, images, annotations)?
• How to tame intra-class
Results: Weakly-supervised detection on PASCAL VOC
Problems with human supervision
• Biased, non-comprehensive
• Concept-specific expertise
• Early hard decisions
Open challenges
Which detection to pick?
What defines a category?
Proposed approach: Webly-supervised learning
Fully automated system: Train your own concept!
237 Concepts
90,000 detectors
Merging synonyms
Detector training
75,000,000 images
18,000,000 annotations