Task-Aware Weakly Supervised Object Localization With Transformer.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised object localization (WSOL) typically uses class-specific regions, leading to incomplete localization.
• Existing methods struggle with accurate object boundary prediction using only image-level labels.

Purpose of the Study:

• To develop an end-to-end framework for accurate and class-agnostic object localization in WSOL.
• To improve both localization and classification performance simultaneously.

Main Methods:

• Proposed a task-aware framework (TAFormer) with a transformer encoder-decoder architecture.
• Implemented a representation encoder, localization decoder, and classification decoder for class-agnostic foreground map generation.
• Utilized an optimal transport algorithm for pixel-level pseudo-labeling to refine foreground maps.

Main Results:

• TAFormer demonstrated remarkable performance in both class-agnostic localization and classification.
• Achieved favorable results against state-of-the-art methods on standard benchmarks.
• Showcased enhanced robustness against adversarial attacks and noisy labels.

Conclusions:

• TAFormer offers a novel approach to WSOL by combining transformer architecture and optimal transport for accurate localization.
• The framework effectively addresses the limitations of class-specific region reliance in prior methods.
• TAFormer presents a robust and high-performing solution for weakly supervised object localization tasks.