Large-Scale Coarse-to-Fine Object Retrieval Ontology and Deep Local Multitask Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object retrieval is crucial for applications like video surveillance and e-commerce.
• Existing systems face challenges with large datasets, imbalanced data, and fine-grained details.
• Deep learning models often struggle with these complexities, necessitating improved approaches.

Purpose of the Study:

• To develop a robust coarse-to-fine object retrieval (CFOR) system for large-scale datasets.
• To enhance object retrieval performance from category to attribute levels.
• To address challenges including imbalanced data, viewpoint variations, and background clutter.

Main Methods:

• Integration of object ontology, a local multitask deep neural network (local MDNN), and an imbalanced data solver.
• Leveraging object ontology for inner-group correlations and computational efficiency.
• Utilizing a local MDNN to link ontology with raw data and an MCC-based solver for data imbalance.

Main Results:

• The CFOR system demonstrates robustness against common object retrieval challenges.
• The local MDNN framework improved recall rate by 14.2% in attribute learning compared to single-task learning (STL).
• The imbalanced data solver boosted recall rate by 5.14% for fewer data attributes, achieving MAP@30 of 0.815.

Conclusions:

• The proposed CFOR system effectively improves large-scale object retrieval performance.
• The synergistic combination of object ontology, local MDNN, and imbalanced data solver is key to the system's success.
• The approach offers a novel solution for attribute learning and retrieval in imbalanced, large-scale datasets.