Background-Aware Classification Activation Map for Weakly Supervised Object Localization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised object localization (WSOL) uses image-level annotations, simplifying data requirements compared to dense annotations.
• Existing WSOL methods often overlook background influences, leading to excessive background activations and inaccurate localization.
• This limitation hinders the precise identification of object boundaries and locations in complex scenes.

Purpose of the Study:

• To propose a novel mechanism, the background-aware classification activation map (B-CAM), to enhance WSOL.
• To introduce background awareness into the WSOL training process to mitigate background interference.
• To improve the accuracy and robustness of object localization using only image-level annotations.

Main Methods:

• Developed B-CAM, a mechanism that incorporates background awareness into WSOL training.
• Aggregated object image-level features for standard supervision.
• Introduced an additional background image-level feature to represent pure-background samples, providing background cues to suppress background activations.
• Trained a background classifier with image-level annotations to generate adaptive background scores for binary localization mask determination.

Main Results:

• Demonstrated the effectiveness of the proposed B-CAM approach.
• Achieved improved object localization performance across diverse WSOL benchmarks.
• Successfully suppressed background activations, leading to more precise object localization maps.

Conclusions:

• The B-CAM mechanism effectively addresses the challenge of background interference in WSOL.
• Integrating background awareness significantly enhances the performance of weakly supervised object localization.
• The proposed method shows strong generalization capabilities across multiple benchmark datasets (CUB-200, ILSVRC, OpenImages, VOC2012).