Interferences in Match Kernels.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Image representations like bag-of-visual-words, Fisher vectors, and VLAD aggregate local descriptors into single vectors.
• Comparing these representations using dot-products results in match kernels, where unrelated descriptors can interfere with similarity scores.

Purpose of the Study:

• To formalize the problem of interference in match kernels for image representation.
• To propose and evaluate two novel methods for equalizing local descriptor contributions in aggregated image vectors.

Main Methods:

• Developed two weighting strategies to modify the aggregation stage of image representations.
• The first method democratizes descriptor importance for set comparison.
• The second method equalizes the match of individual descriptors to the aggregated vector.

Main Results:

• Both proposed methods significantly reduce interference in match kernels.
• Achieved substantial performance improvements over state-of-the-art methods in image search.
• Demonstrated effectiveness on standard public image retrieval benchmarks with short and mid-size vectors.

Conclusions:

• Equalizing local descriptor contributions through learned weights enhances image representation quality.
• The proposed methods offer a significant performance boost for image retrieval systems.
• These techniques are particularly effective for efficient image search using compact vector representations.