Central Similarity Quantization for Efficient Image and Video Retrieval (1908.00347v5)

Abstract: Existing data-dependent hashing methods usually learn hash functions from pairwise or triplet data relationships, which only capture the data similarity locally, and often suffer from low learning efficiency and low collision rate. In this work, we propose a new \emph{global} similarity metric, termed as \emph{central similarity}, with which the hash codes of similar data pairs are encouraged to approach a common center and those for dissimilar pairs to converge to different centers, to improve hash learning efficiency and retrieval accuracy. We principally formulate the computation of the proposed central similarity metric by introducing a new concept, i.e., \emph{hash center} that refers to a set of data points scattered in the Hamming space with a sufficient mutual distance between each other. We then provide an efficient method to construct well separated hash centers by leveraging the Hadamard matrix and Bernoulli distributions. Finally, we propose the Central Similarity Quantization (CSQ) that optimizes the central similarity between data points w.r.t.\ their hash centers instead of optimizing the local similarity. CSQ is generic and applicable to both image and video hashing scenarios. Extensive experiments on large-scale image and video retrieval tasks demonstrate that CSQ can generate cohesive hash codes for similar data pairs and dispersed hash codes for dissimilar pairs, achieving a noticeable boost in retrieval performance, i.e. 3\%-20\% in mAP over the previous state-of-the-arts. The code is at: \url{https://github.com/yuanli2333/Hadamard-Matrix-for-hashing}

• The paper presents CSQ, a method that refines hash functions by clustering similar pairs around common centers for efficient retrieval.
• It employs Hadamard matrices and Bernoulli distributions to compute hash centers with ensured minimum distances in the Hamming space.
• Benchmark results indicate CSQ improves mean Average Precision by 3% to 20% in large-scale image and video retrieval tasks.

Central Similarity Quantization for Efficient Image and Video Retrieval

The paper "Central Similarity Quantization for Efficient Image and Video Retrieval" introduces an innovative approach to hashing, tackling inherent inefficiencies in traditional methods that primarily rely on local similarities. The authors propose a method named Central Similarity Quantization (CSQ), which refines the way hash functions are learned, significantly improving both the efficiency of learning and retrieval accuracy.

Core Contributions and Methodology

CSQ introduces a global similarity concept known as central similarity. Traditional hashing methods often use pairwise or triplet data to determine similarities, which may lead to low learning efficiency and suboptimal retrieval performance due to their localized perspective. CSQ, in contrast, encourages hash codes for similar data pairs to gather around a common center, while dissimilar pairs should gravitate toward different centers.

The key innovation lies in the definition and computation of "hash centers," which are strategically distributed points in the Hamming space designed to maintain a minimum distance. These centers are derived in an efficient manner using the Hadamard matrix and Bernoulli distributions, facilitating better separation and allocation in hash space. By doing so, CSQ achieves a reduction in time complexity and enhances the discrimination capability of hash codes by capturing global data distributions.

Numerical Results

The implementation of CSQ shows robust performance improvements across several benchmarks. For instance, in large-scale image and video retrieval scenarios, it reports a noticeable improvement in mean Average Precision (mAP) of 3%-20% over prior methods. Such results emphasize the potential of CSQ in providing more cohesive hash codes for similar data and more dispersed codes for dissimilar ones, leveraging global material relations for enhanced retrieval accuracy.

Implications and Future Directions

Practically, CSQ offers a scalable, efficient solution for vast imagery and video datasets, particularly beneficial in data-intensive industries where rapid and accurate data retrieval is paramount. Theoretically, it introduces a shift from localized to global similarity considerations in hashing, potentially inspiring further refinements in the data retrieval domain.

Looking forward, the paper suggests exploring learned hash centers, as opposed to the pre-computed centers, for potentially more flexible applications. Future work might delve into dynamically adjusting these centers based on data evolution or leveraging deeper insights from the data itself, such as intrinsic data structures or real-time adaptability.

Conclusion

Central Similarity Quantization represents a substantial step forward in hashing methodologies for retrieval tasks. By capitalizing on global similarities and efficiently structuring the hash space, this method enhances both the speed and precision of retrieval processes. As the field of artificial intelligence continues to progress, methodologies like CSQ could lay the groundwork for increasingly adaptive and intelligent systems, capable of seamless integration across varied data domains.