Module of Axis-based Nexus Attention for weakly supervised object localization.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Weakly supervised object localization (WSOL) struggles to segment entire objects, often focusing only on discriminative parts.
• Existing corruption-based WSOL methods require a hyperparameter for corruption degree and yield coarse localization.
• These methods can disrupt training and fail to capture fine-grained object details.

Purpose of the Study:

• To introduce a novel approach, Module of Axis-based Nexus Attention (MoANA), for improved weakly supervised object localization.
• To address the challenge of segmenting entire objects by adaptively activating less discriminative regions.
• To eliminate the need for additional hyperparameters like corruption thresholds in WSOL.

Main Methods:

• MoANA utilizes a triple-view attention representation, attention expansion, and a feature calibration mechanism.
• It trains fine-grained attention maps with element-wise attention values, unlike methods with uniform attention.
• The approach adaptively activates both discriminative and less discriminative object regions.

Main Results:

• MoANA demonstrates superior performance compared to existing weakly supervised object localization methods.
• The proposed method achieves more elaborate and accurate localization of entire objects.
• Analysis of MoANA's components and visualization of attention maps provide insights into its effectiveness.

Conclusions:

• MoANA offers a robust and hyperparameter-free solution for weakly supervised object localization.
• The method effectively localizes entire objects by integrating discriminative and non-discriminative regions.
• MoANA advances the state-of-the-art in attention-based object localization techniques.