Pixel-Level Domain Adaptation: A New Perspective for Enhancing Weakly Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly Supervised Semantic Segmentation (WSSS) learns from image-level tags, but current methods using Class Activation Maps (CAMs) suffer from imbalanced activation, locating only discriminative object parts.
• This imbalanced activation leads to incomplete and imprecise pseudo-masks, hindering the performance of WSSS models.
• The distribution discrepancy between discriminative and non-discriminative object parts is a key challenge in generating accurate pseudo-masks.

Purpose of the Study:

• To propose a novel Pixel-Level Domain Adaptation (PLDA) method for WSSS.
• To address the imbalanced activation issue in WSSS by learning pixel-wise domain-invariant features.
• To improve the completeness and precision of pseudo-masks generated for WSSS.

Main Methods:

• Introduced a Pixel-Level Domain Adaptation (PLDA) method for WSSS.
• Employed adversarial training with a multi-head domain classifier to learn pixel-wise domain-invariant features.
• Developed a Confident Pseudo-Supervision strategy to enhance pixel discriminability for segmentation.

Main Results:

• The proposed PLDA method effectively addresses the imbalanced activation issue in WSSS.
• Experimental results demonstrate significant improvements when integrating the PLDA method with existing WSSS approaches.
• The approach leads to more complete and precise pseudo-mask generation.

Conclusions:

• The PLDA method offers a simple yet effective solution for improving WSSS by learning domain-invariant pixel features.
• The Confident Pseudo-Supervision strategy complements adversarial training, enhancing segmentation performance.
• The proposed method is intuitive and easily integrable into current WSSS frameworks, showing broad applicability.