Adversarial Transformers for Weakly Supervised Object Localization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised object localization (WSOL) uses image-level labels for object detection, offering practical advantages over fully supervised methods.
• Existing WSOL techniques struggle with generating precise localization maps due to the lack of pixel-level supervision, limiting overall performance.
• The challenge lies in refining localization accuracy without detailed ground truth annotations.

Purpose of the Study:

• To develop a novel weakly supervised object localization method that overcomes the limitations of imprecise localization maps.
• To introduce an adversarial transformer network (ATNet) capable of generating accurate pixel-level pseudo labels for improved localization.
• To enhance the scalability and practicability of object localization through advanced deep learning techniques.

Main Methods:

• Proposing an Adversarial Transformer Network (ATNet) comprising an object transformer (G) and a part transformer (D).
• Utilizing an adversarial training process where G generates localization maps and pseudo labels, while D refines discrimination of localization details.
• Employing transformers with adversarial training for the first time in WSOL to achieve a well-learned localization model.

Main Results:

• ATNet demonstrated favorable performance against state-of-the-art WSOL methods across two standard benchmarks using four different backbones.
• The proposed method effectively generates accurate localization maps and pixel-level pseudo labels.
• Adversarial training within ATNet enhanced the model's robustness against adversarial attacks.

Conclusions:

• ATNet successfully addresses the challenge of rough localization maps in WSOL by leveraging adversarial transformers and pseudo labels.
• The proposed approach achieves superior performance and robustness compared to existing WSOL methods.
• This work pioneers the use of transformers combined with adversarial training for effective weakly supervised object localization.