Large-Scale Cross-Modality Search via Collective Matrix Factorization Hashing.

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• Hashing transforms data into binary representations for efficient search and storage.
• Large-scale cross-modality search is crucial due to the proliferation of multimodal data on the web.
• Existing methods struggle with generating effective hash codes for diverse data types.

Purpose of the Study:

• To propose a novel framework for multimodal hashing called Collective Matrix Factorization Hashing (CMFH).
• To enable accurate large-scale cross-modality search by learning unified hash codes for different data modalities.
• To extend the framework for both unsupervised and supervised learning scenarios.

Main Methods:

• Collective Matrix Factorization Hashing (CMFH) learns unified hash codes in a shared latent semantic space.
• The framework connects different modalities of a single instance for effective retrieval.
• Extensions include unsupervised methods preserving Euclidean structure and supervised methods using label information.

Main Results:

• CMFH was evaluated on three benchmark datasets for cross-modality search.
• The proposed method demonstrated significant performance improvements over state-of-the-art approaches.
• Experimental results validate the effectiveness of CMFH in multimodal retrieval.

Conclusions:

• CMFH provides an effective solution for multimodal hashing and cross-modality search.
• The unified latent semantic space facilitates accurate retrieval across different data types.
• The framework's adaptability to both supervised and unsupervised learning enhances its applicability.