Effective Multi-Query Expansions: Collaborative Deep Networks for Robust Landmark Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Information Retrieval

Background:

• Existing landmark retrieval methods primarily rely on landmark geometry, which can be unreliable due to varying user viewpoints and photo quality.
• Low-quality landmark shapes from user-generated content present significant challenges and are understudied in current research.

Purpose of the Study:

• To develop a novel framework for semantically robust landmark retrieval that overcomes limitations of existing geometric approaches.
• To improve the accuracy and reliability of landmark retrieval from diverse, potentially low-quality social media images.

Main Methods:

• A multi-query expansion technique is proposed, first identifying top-k photos based on latent topics to create a robust multi-query set.
• Latent Dirichlet Allocation (LDA) techniques are extended to identify latent topics for landmark photos.
• Collaborative deep networks are learned using matrix factorization on a user-photo matrix derived from the multi-query set to extract high-level features.

Main Results:

• The proposed framework generates high-level semantic features for landmark photos, enabling more robust similarity matching.
• The multi-query set is compacted within the learned feature space for efficient retrieval.
• Experimental results on real-world social media data demonstrate superior performance compared to existing landmark retrieval methods.

Conclusions:

• The multi-query expansion framework effectively retrieves semantically robust landmarks, even from low-quality images.
• Collaborative deep networks provide a powerful approach for learning discriminative features in landmark retrieval.
• This research offers a significant advancement in landmark retrieval from user-generated visual content.