Paper Site: https://www1.cs.columbia.edu/CAVE/publications/pdfs/Kumar_ECCV08.pdf
Problem Definition
Using a image search engine based entirely on face, search through a large scale database automatically labeled on the basis of several facial attributes with simple text queries.
Contribution and Discussion
Created the first face search engine allowing users to search through large collections of images which have been autmatically labeled.
Using a novel combination of Support Vector Machines and Adaboost shows srate-of-the-art classification results compared to previous works.
Fully automatic, easy to scale and computes all labels off-line, leading to fast on-line search performance.
Methods
- Build a large database of over 3.1 million face images extracted from over 6.2 million images collected from the internet.
Overview of database creation
- Create a rich set of local feature options from which the classifier can automatically select the best ones: the region of the face to extract features from, the type of pixel data to use, the kind of normalization to apply to the data, and finally, the level of aggregation to use.
The face regions used for automatic feature selection
- Combine the strengths of these two methods by constructing a number of “l(fā)ocal” SVMs and letting Adaboost create an optimal classifier using a linear combination of them. We create one SVM for each region, feature type, and SVM parameter combination, using the LibSVM library.
