A Componentwise Approach to Weakly Supervised Semantic Segmentation Using Dual-Feedback Network.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised semantic segmentation methods rely on pseudolabels to infer spatial information.
• Existing pseudolabels often suffer from mislabeled regions and inaccurate boundaries, limiting segmentation performance.
• Incomplete recovery of position information is a key challenge in current approaches.

Purpose of the Study:

• To develop a componentwise approach for recovering high-level semantic and low-level physical information independently.
• To propose a novel dual-feedback network (DFN) for improved weakly supervised semantic segmentation.
• To enhance the accuracy of pseudolabels by correcting mislabeled regions and refining object boundaries.

Main Methods:

• Decomposition of position information into semantic and physical components.
• Iterative pseudolabel updating mechanism to refine high-level semantic information.
• Superpixel-based random walk mechanism for reconstructing low-level physical information.
• Integration of these mechanisms within a novel dual-feedback network (DFN) architecture.

Main Results:

• The proposed DFN effectively corrects mislabeled regions and refines object boundaries in pseudolabels.
• Experiments demonstrate superior performance compared to state-of-the-art methods.
• Significant improvements in mean intersection-over-union (mIoU) were achieved on benchmark datasets.

Conclusions:

• The componentwise approach effectively recovers lost position information in weakly supervised semantic segmentation.
• The dual-feedback network (DFN) offers a robust solution for generating accurate pseudolabels.
• DFN advances the state-of-the-art in weakly supervised semantic segmentation, achieving higher mIoU.