CoReS: Compatible Representations via Stationarity.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Visual search systems require feature extraction from large datasets.
• Updating representation models necessitates re-indexing, which is computationally expensive and sometimes infeasible.
• Current methods lack compatibility between old and new features, hindering model upgrades.

Purpose of the Study:

• To propose a novel training procedure for learning compatible feature representations.
• To enable interchangeable use of old and new features over time in visual search.
• To eliminate the need for re-indexing gallery sets during model updates.

Main Methods:

• Introduced CoReS, a training procedure for compatible representations.
• Utilized fixed classifiers based on polytopes for feature stationarity.
• Ensured maximal class separation and stationary spatial configuration in the representation space.

Main Results:

• The proposed CoReS method demonstrates superior performance over existing state-of-the-art techniques.
• The procedure is highly effective for multiple training set upgrades, common in real-world applications.
• Eliminates the need for mapping or pairwise training between old and new models.

Conclusions:

• CoReS offers a computationally efficient and practical solution for updating visual search models.
• The method ensures feature compatibility, enabling seamless integration of new data without re-indexing.
• This advancement is particularly valuable for large-scale visual search applications with evolving datasets.