Florian Schroff, Antonio Criminisi & Andrew ZissermanICCV 2007

Harvesting Image Databases from the Web

OutlineGoal: 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

1. enter object keyword (e.g. penguin)

2. retrieve set of images using Google web search

3. filter to remove drawings & abstract images

4. rank images using meta-data from web pages

5. train SVM on visual features using (4) as noisy training data

6. final ranking using trained SVM

Example: Penguin

1. Enter “penguin”

2. 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

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

Details of Abstract Filter

Details of Meta-data Re-rank Filter

Example: Penguin continued

More examples classes – cars, elephants

More examples classes – watches, zebras