Asymmetric distances for binary embeddings.

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• Binary embedding of image signatures enhances data compression and search efficiency in large-scale retrieval.
• Traditional methods often binarize both query and database signatures symmetrically.

Purpose of the Study:

• To propose novel asymmetric distance measures for binary embedding techniques.
• To evaluate the effectiveness of asymmetric schemes in improving retrieval accuracy compared to symmetric methods.

Main Methods:

• Developed two general asymmetric distance metrics applicable to various embedding algorithms.
• Tested the proposed methods on four public benchmarks with up to 1 million images.
• Compared performance against the standard symmetric Hamming distance.

Main Results:

• Asymmetric distances consistently yielded significant accuracy improvements across all tested binary embedding techniques.
• The proposed methods demonstrated superior performance over symmetric Hamming distance on large-scale datasets.
• The benefits of data compression and search efficiency were maintained.

Conclusions:

• Asymmetric embedding schemes offer a more accurate approach to large-scale query-by-example image retrieval.
• The proposed asymmetric distances provide a general and effective enhancement for existing binary embedding methods.
• This work highlights the advantage of not binarizing query signatures in certain retrieval scenarios.