Learning image similarity from Flickr groups using fast kernel machines

Area of Science:

• Computer Vision
• Machine Learning
• Image Analysis

Background:

• Measuring image similarity is crucial in computer vision.
• Existing methods often rely on conventional visual features, which may not capture nuanced similarities.
• Online image communities like Flickr offer rich, diverse data for learning image relationships.

Purpose of the Study:

• To develop a novel method for measuring image similarity by leveraging data from online Flickr image groups.
• To introduce an efficient training algorithm for Support Vector Machine (SVM) classifiers used in this similarity measure.
• To evaluate the proposed approach's performance against conventional methods in image matching, retrieval, and classification tasks.

Main Methods:

• Utilized 103 Flickr groups to train one-versus-all multiclass classifiers.
• Employed SVM with a histogram intersection kernel for classification.
• Developed the Stochastic Intersection Kernel MAchine (SIKMA) training algorithm for rapid training on large datasets.
• Computed image similarity based on the distance between feature vectors derived from classifier responses.

Main Results:

• The proposed image similarity measure demonstrated superior performance on the Corel and PASCAL VOC 2007 datasets.
• Achieved better results in image matching, retrieval, and classification compared to conventional visual features.
• The SIKMA algorithm enabled accurate SVM training on tens of thousands of examples within minutes, outperforming linear classifiers.

Conclusions:

• Learning image similarity from online image groups provides a robust and effective approach.
• The SIKMA algorithm significantly improves the efficiency and scalability of training kernel classifiers.
• This method offers a promising alternative for various computer vision applications requiring accurate image similarity assessment.