Reproduce “Codebook-Centric Deep Hashing: End-to-End Joint Learning of Semantic Hash Centers and Neural Hash Function” on OpenResearch

About this paper

Shuo Yin, Zhiyuan Yin, Yuqing Hou, Rui Liu, Yong Chen, Dell Zhang

The Center-Reassigned Hashing (CRH) framework jointly learns semantic hash centers and a neural hash function in an end-to-end manner, dynamically reassigning centers from a codebook. It achieves state-of-the-art retrieval performance across image and video datasets, demonstrating mAP improvements such as 2.1%-2.6% on Stanford Cars and 4.8%-6.6% on NABirds.

Key insights

• Dynamic and iterative reassignment of hash centers in conjunction with hash function optimization enables the end-to-end learning of semantically meaningful binary codes that adapt to evolving data representations.
• Learning hash centers from a flexible, pre-defined codebook with a greedy reassignment algorithm provides an efficient and effective approximation to the complex problem of finding optimal hash center assignments.
• The multi-head architecture effectively expands the semantic expressiveness of hash centers, improving their ability to model intricate inter-class relationships and directly correlating with enhanced retrieval performance.

Summary by alphaXiv.