Reproduce “Central Similarity Quantization for Efficient Image and Video Retrieval” on OpenResearch

About this paper

Li Yuan, Tao Wang, Xiaopeng Zhang, Francis EH Tay, Zequn Jie, Wei Liu, Jiashi Feng

Central Similarity Quantization (CSQ) redefines deep hashing by leveraging pre-computed "hash centers" and a novel "central similarity" metric to efficiently learn compact binary codes for image and video retrieval. This approach achieves state-of-the-art retrieval accuracy, with mAP increases of up to 20% on ImageNet, while also demonstrating significantly faster training, being 3 to 5.5 times quicker than previous methods.

Key insights

• Modeling global data similarity through well-defined hash centers significantly improves hash code discriminability and retrieval performance compared to methods based on local pairwise or triplet relationships.
• Pre-computed binary hash centers with guaranteed separation in Hamming space are more effective and computationally trivial than learning non-binary centers from data features and subsequently binarizing them.
• The central similarity loss adequately captures complex data relationships, rendering the inclusion of traditional local pairwise similarity losses largely redundant within this framework.

Summary by alphaXiv.