Simultaneously Discovering and Localizing Common Objects in Wild Images.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Supervised and weakly supervised methods have advanced common object discovery.
• Traditional object localization requires extensive annotations (bounding boxes or image-level labels).
• A fully unsupervised approach is needed for object discovery without any human-provided labels.

Purpose of the Study:

• To develop a fully unsupervised method for simultaneous common object discovery and localization.
• To address the challenge of discovering objects in images without any annotations.
• To improve upon existing methods by handling complex image scenarios and enhancing image retrieval.

Main Methods:

• Formulated unsupervised object discovery as a sub-graph mining problem on a weighted graph of object proposals.
• Utilized a maximal-flow-based algorithm for efficient optimization.
• Developed a method that jointly discovers positive images and common objects by identifying strongly connected sub-graphs representing object patterns.

Main Results:

• Successfully discovered and localized common objects across various classes without supervision.
• Demonstrated robustness to scale, viewpoint, appearance variations, and partial occlusions.
• Achieved significant improvements in image retrieval tasks by considering inter-image similarities.

Conclusions:

• The proposed fully unsupervised method effectively performs common object discovery and localization.
• The approach is adaptable to image retrieval, outperforming existing methods.
• This work advances the field by enabling object discovery in scenarios lacking any annotations.