Reproduce “Unsupervised Hashing with Similarity Distribution Calibration” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kam Woh Ng, Xiatian Zhu, Jiun Tian Hoe, Chee Seng Chan, Tianyu Zhang, Yi-Zhe Song, Tao Xiang
A University of Surrey study introduces Similarity Distribution Calibration (SDC), an approach addressing the similarity collapse issue in unsupervised hashing for image retrieval. SDC achieves higher mean Average Precision (mAP) across various benchmarks, for instance reaching 89.0 mAP on CIFAR10 (64-bit, ViT-B/16) and 63.8 mAP on RParis (512-bit), by regularizing the distribution of hash code similarities.
Key insights
- The paper identifies similarity collapse as a fundamental, largely overlooked flaw in existing similarity preservation-based unsupervised hashing methods.
- Calibrating the distribution of hash code similarities to a well-spread target distribution (e.g., Beta distribution) effectively counteracts the inherent biases and limited capacity of the Hamming space.
- SDC provides a principled, modular solution that enhances the discriminative power of learned hash codes and improves overall retrieval performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.