Abstract
Learning effective binary codes for image retrieval typically requires large datasets, which are often unavailable for specific target domains. This paper presents Optimal Projection Guided Transfer Hashing (GTH), an unsupervised framework that leverages knowledge from a data-rich source domain to improve retrieval in a data-scarce target domain. By aligning hashing projections through robust maximum likelihood estimation (MLE), GTH mitigates the domain gap and achieves state-of-the-art performance without requiring labeled data.
Introduction
In large-scale image retrieval, hashing compresses high-dimensional images into compact binary strings, enabling rapid search and reduced storage. However, modern "learning to hash" (L2H) algorithms are data-hungry, requiring thousands of images to learn precise models. In many real-world scenarios, we encounter a "small data" problem in the target domain of interest. While general-purpose source domains (e.g., web images) may be abundant, a "domain gap"—differences in lighting, camera angles, or composition—often prevents the direct application of source-trained models. GTH aims to bridge this gap by transferring knowledge at the projection level rather than the data level.
Methodology
The core innovation of GTH is its robust approach to aligning hashing functions across domains. The method consists of several key components:
- Dual Projections: GTH learns two separate hashing functions: $W_t$ for the target domain and $W_s$ for the source domain.
- Robust Maximum Likelihood Estimation (MLE): Traditional transfer methods often use simple $l_2$ distances to align models, assuming Gaussian noise. GTH proposes an MLE-based solution using an iteratively weighted $l_2$ loss. This sigmoid-based weighting effectively "down-weights" large discrepancies between source and target projections, making the transfer resilient to outliers or irrelevant source features.
- Alternating Optimization: The model is trained by iteratively updating the target projection, the source projection, and the binary codes. To ensure the projections remain orthogonal and mathematically sound, the authors employ the Crank-Nicolson scheme.
- Unsupervised Learning: Crucially, GTH requires no manual labels for either domain, making it highly applicable to "in-the-wild" datasets.
Experimental Results
The authors evaluated GTH on three major benchmarks: PIE (face images), Office (Amazon & Dslr daily objects), and VLCS (VOC2007 & Caltech101 photographic datasets). GTH was compared against state-of-the-art baselines including LSH, ITQ, and OCH.
Key Findings
- Superior Accuracy: The robust version, GTH-h, consistently outperformed all baselines across all datasets and code lengths (16 to 64 bits). For instance, on the PIE dataset with 16-bit codes, GTH-h achieved a Mean Average Precision (MAP) of 25.45%, compared to 21.06% for the best baseline (NoDA).
- Robustness to Small Data: GTH maintains high performance even when the number of target training images is significantly reduced, demonstrating its effectiveness in data-scarce scenarios.
- Scalability: As code length increases, GTH-h maintains its performance lead, reaching 63.53% MAP on the challenging VOC2007 & Caltech101 benchmark at 64 bits.
Conclusion and Future Work
GTH demonstrates that aligning hashing functions via robust MLE is a powerful strategy for unsupervised transfer hashing. By focusing on projection-level similarity and weighting out noise, it overcomes the limitations of data-hungry L2H methods. Future work may explore extending this framework to non-linear projections or integrating it into end-to-end deep learning architectures to capture more complex semantic features.
Summary updated