Florian Schroff, Antonio Criminisi & Andrew Zisserman ICCV 2007

1. Florian Schroff, Antonio Criminisi & Andrew ZissermanICCV 2007 • Harvesting Image Databases from the Web

2. Outline Goal: retrieve class specific images from the web • images are ranked using a multi-modal approach: • text & meta data from the web pages • visual features Algorithm enter object keyword (e.g. penguin) retrieve set of images using Google web search filter to remove drawings & abstract images rank images using meta-data from web pages train SVM on visual features using (4) as noisy training data final ranking using trained SVM

3. Example: Penguin • Enter “penguin” • Retrieve images from web pages returned by Google web search on penguin • 522 in-class, 1771 non-class • 3. Remove drawings & abstract images • 391 in-class, 784 non-class

4. Example: Penguin continued 4. rank images using naïve Bayes metadata ranker 5. Train SVM on visual features using ranked images as noisy training data 6. Final re-ranking using trained SVM

5. Details of Abstract Filter

6. Details of Meta-data Re-rank Filter

7. Example: Penguin continued

8. More examples classes – cars, elephants

9. More examples classes – watches, zebras