Asymmetric Transfer Hashing With Adaptive Bipartite Graph Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Learning to hash is crucial for visual retrieval due to efficiency.
• Existing methods fail in heterogeneous cross-domain retrieval.
• Cross-domain retrieval faces domain and feature gaps.

Purpose of the Study:

• Propose a generalized image transfer retrieval (GITR) problem.
• Introduce an asymmetric transfer hashing (ATH) framework to address GITR.
• Develop unsupervised, semi-supervised, and supervised versions of ATH.

Main Methods:

• ATH characterizes domain gaps via asymmetric hash functions.
• Minimizes feature gaps using an adaptive bipartite graph.
• Preserves single-domain data structure with a domain affinity graph.

Main Results:

• ATH achieves effective knowledge transfer across domains.
• Avoids information loss during feature alignment.
• Demonstrates superior performance over state-of-the-art methods.

Conclusions:

• The proposed ATH framework effectively addresses the challenges of GITR.
• ATH offers a robust solution for heterogeneous cross-domain visual retrieval.
• The method shows significant improvements on various benchmarks.