Dataset-Driven Unsupervised Object Discovery for Region-Based Instance Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Instance image retrieval benefits from object discovery for better feature representation and user guidance.
• Object classes are typically undefined, and labels are unavailable in image retrieval datasets, posing a challenge for object discovery.

Purpose of the Study:

• To propose a novel dataset-driven unsupervised object discovery framework for image retrieval.
• To address the challenge of object discovery in datasets lacking predefined object classes and labels.

Main Methods:

• Utilizing deep feature representation and weakly-supervised object detection to derive supervisory information from within the dataset.
• Constructing class-wise object detectors and assigning multiple detectors per image.
• Introducing a "base-detector repository" for efficient object detector construction in large datasets.
• Implementing a self-boosting mechanism for iterative refinement of object discovery.

Main Results:

• The proposed framework achieves more accurate object discovery compared to existing unsupervised methods.
• The framework successfully generates object discovery results without manual annotation or auxiliary datasets.
• Demonstrated improved performance for region-based instance image retrieval.

Conclusions:

• The novel framework effectively addresses unsupervised object discovery in image retrieval.
• The method enhances instance image retrieval performance by improving object detection accuracy.
• This approach offers a viable alternative to supervised methods, eliminating the need for manual data labeling.