Mining compact bag-of-patterns for low bit rate mobile visual search

Area of Science:

• Computer Vision
• Image Recognition
• Mobile Computing

Background:

• Existing visual pattern methods rely on 2D image data, leading to inaccuracies due to depth variations.
• Designing compact descriptors for these patterns remains a challenge, hindering efficient image retrieval.

Purpose of the Study:

• To propose a novel compact bag-of-patterns (CBoPs) descriptor for improved mobile landmark search.
• To address the limitations of 2D visual patterns by incorporating 3D information.
• To develop a descriptor that enables low bit rate mobile landmark search with reduced latency.

Main Methods:

• Constructing 3D point clouds from reference images to capture real-world concurrences of visual words.
• Employing a novel gravity distance metric for mining discriminative visual patterns.
• Utilizing sparse coding over mined patterns to create the CBoPs descriptor, optimizing for reconstruction and coding length.

Main Results:

• The CBoPs descriptor achieves comparable accuracy to million-scale bag-of-words histograms.
• Demonstrated a high descriptor compression rate (approximately 100 bits), significantly outperforming existing schemes.
• A low bit rate mobile landmark search prototype utilizing CBoPs showed reduced query delivery latency.

Conclusions:

• The proposed CBoPs descriptor effectively overcomes the limitations of 2D visual patterns by leveraging 3D data.
• CBoPs offer a highly compressed yet accurate representation for efficient mobile landmark search applications.
• This approach advances the field of compact image descriptors for resource-constrained mobile environments.