Boosting Weakly Supervised Object Localization and Segmentation With Domain Adaption.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised object localization (WSOL) uses image-level tags for pixel-level localization, reducing annotation effort.
• Existing one-stage WSOL methods often focus on discriminative object parts due to multi-instance learning limitations.
• A significant challenge in WSOL is the domain shift between training and testing data.

Purpose of the Study:

• To reframe weakly supervised object localization (WSOL) as a domain adaptation (DA) task.
• To develop a novel DA-WSOL pipeline addressing domain shift specificities in WSOL.
• To improve the localization of entire objects rather than just discriminative parts.

Main Methods:

• Proposed a domain adaptation framework (DA-WSOL) tailored for WSOL challenges.
• Introduced a target sampling strategy to identify source-related and Universum samples.
• Addressed sample unbalancing and label mismatch issues between source and target domains.

Main Results:

• The DA-WSOL pipeline demonstrated superior performance compared to state-of-the-art (SOTA) methods on three WSOL benchmarks.
• The proposed method effectively localized entire objects, overcoming limitations of previous approaches.
• Significant improvements were observed in downstream weakly supervised semantic segmentation tasks.

Conclusions:

• Viewing WSOL as a domain adaptation task offers a promising new perspective.
• The developed DA-WSOL pipeline effectively mitigates domain shift issues in WSOL.
• This approach enhances both object localization accuracy and the performance of related computer vision tasks.