Anti-Adversarially Manipulated Attributions for Weakly Supervised Semantic Segmentation and Object Localization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Accurate pixel-level localization is essential for weakly supervised semantic segmentation and object localization.
• Existing attribution maps often focus on small, discriminative object regions, neglecting broader class-relevant features.

Purpose of the Study:

• To develop a method that generates more comprehensive pixel-level localization maps for target objects.
• To improve performance in weakly supervised semantic segmentation and object localization tasks.

Main Methods:

• Introduced AdvCAM (Attribution Map using Class Activation Maps), which manipulates images in an anti-adversarial manner to enhance non-discriminative features.
• Utilized pixel gradients opposite to adversarial attack directions to perturb images and improve feature attribution.
• Implemented a novel regularization technique to prevent incorrect attributions to unrelated regions and over-concentration on small object parts.

Main Results:

• AdvCAM achieved state-of-the-art performance in weakly and semi-supervised semantic segmentation on PASCAL VOC 2012 and MS COCO 2014 datasets.
• The method also set new state-of-the-art benchmarks in weakly supervised object localization on CUB-200-2011 and ImageNet-1K datasets.
• Enhanced attribution maps successfully identified more regions of the target object compared to previous methods.

Conclusions:

• AdvCAM effectively enhances attribution maps for improved object localization by leveraging anti-adversarial manipulation.
• The proposed regularization further refines attribution maps, leading to superior performance in segmentation and localization tasks.
• This work advances the capabilities of weakly supervised learning for detailed image understanding.