Hashing with Angular Reconstructive Embeddings.

Area of Science:

• Computer Vision
• Machine Learning
• Information Retrieval

Background:

• Hashing-based approximate nearest neighbor search is vital for large-scale visual analysis.
• Existing methods often rely on Euclidean distance, limiting applicability for cosine similarity-based tasks.
• Efficient storage and retrieval are key challenges in content-based visual analysis.

Purpose of the Study:

• To develop a novel hashing method for learning compact binary codes based on cosine similarity.
• To address the limitations of Euclidean distance-based hashing in visual search applications.
• To propose Angular Reconstructive Embeddings (ARE) for improved similarity measurement.

Main Methods:

• Proposed the Angular Reconstructive Embeddings (ARE) method.
• ARE learns binary codes by minimizing reconstruction error between original feature cosine similarities and binary embeddings.
• Developed two efficient algorithms for continuous and discrete optimization of ARE.

Main Results:

• ARE demonstrated superior performance on large-scale image benchmarks.
• The method effectively learns compact binary codes using cosine similarity.
• Outperformed several state-of-the-art hashing techniques in retrieval efficiency and accuracy.

Conclusions:

• Angular Reconstructive Embeddings (ARE) offer an effective solution for learning binary codes with cosine similarity.
• The proposed method enhances large-scale visual search capabilities.
• ARE provides a valuable advancement in hashing-based approximate nearest neighbor search.