Spherical hashing: binary code embedding with hyperspheres.

Area of Science:

• Computer Vision
• Machine Learning
• Data Mining

Background:

• Binary code embedding is crucial for efficient similarity search and compact data representation in large-scale databases.
• Current methods often use hyperplane-based hashing, which may not optimally capture spatial coherence.
• There is a need for advanced embedding techniques that enhance data representation and search efficiency.

Purpose of the Study:

• To introduce a novel hypersphere-based hashing function, termed spherical hashing, for binary code embedding.
• To develop a new binary code distance function, spherical Hamming distance, optimized for the hypersphere-based scheme.
• To generalize spherical hashing to support various kernel-defined similarity measures.

Main Methods:

• Proposed a novel hypersphere-based hashing function (spherical hashing).
• Introduced a tailored binary code distance function (spherical Hamming distance).
• Designed an iterative optimization process for balanced partitioning and hash function independence, supporting kernel functions.

Main Results:

• Spherical hashing significantly outperforms state-of-the-art hyperplane-based methods.
• Achieved up to 100% performance improvement over the second-best method on benchmarks up to 75 million descriptors.
• Demonstrated consistent and large performance gains across GIST, BoW, and VLAD descriptors.

Conclusions:

• Hypersphere-based encoding offers unique advantages for proximity regions in high-dimensional spaces.
• Spherical hashing provides a more spatially coherent data mapping than hyperplane methods.
• The proposed method is intuitive, easy to implement, and offers substantial performance benefits.