TrustArc Webinar - How to Build Consumer Trust Through Data Privacy
Callisto: Content Based Tag Recommendation for Images
1. Callisto – A Content-based Tag
Recommendation Tool
M. Lux, A. Pitman, and O. Marques
2. What does Callisto do?
• Given an image and one or more start tags
• Callisto finds ranked tag recommendations
1. Based on our model (NCP)
2. Based on statistical analysis (Stat)
3. What are the benefits of NCP & Callisto?
• Different tags are suggested.
• Tags are re-ranked based on visual content.
– Consequently:
• With the NCP model, it is common to see tags that are
highly related to visual features being suggested if such
features are there, and not suggested if those features
are missing.
– E.g.: sunset is not suggested if typical colors of sunsets are
missing in the image.
4. The Application
Image to be tagged
Start tag(s)
Low-level
features used
Suggestions
9. Use Case: Juggling
Start tag: juggling
• NCP ranks fire first
• NCP doesn‘t include balls in
the list, which is good, since
there are no balls involved
10. Use Case: Juggling
Start tag: juggling
• NCP suggests portrait
and people
• NCP doesn‘t suggest fire
11. Use Case: Juggling girl
Start tags: juggling girl
• NCP suggests woman
• NCP ranks people higher
12. Performance issues
• Callisto has to download images and tags for
suggestions, which is slow.
• Callisto caches downloads, so next time (with
the same start tag) it is much faster.
• The number of downloaded photos is critical.
– 28 works fine and is not too slow
– 100 is much better, but downloading takes forever