Group-Wise Learning for Weakly Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models require extensive ground-truth data, posing a challenge for tasks like semantic segmentation needing pixel-level annotations.
• Weakly supervised semantic segmentation (WSSS) aims to reduce annotation burden by using less precise labels.

Purpose of the Study:

• To develop a novel group-wise learning framework for WSSS that bridges the gap between image-level annotations and pixel-level segmentation.
• To improve the accuracy and reliability of pseudo ground-truths for training segmentation models.

Main Methods:

• Proposed a group-wise learning framework for WSSS, encoding semantic dependencies across image groups.
• Utilized a graph neural network (GNN) where images are nodes and relations are edges for iterative semantic reasoning.
• Introduced a graph dropout layer to prevent over-reliance on common semantics and enhance object response detection.

Main Results:

• Achieved state-of-the-art performance on PASCAL VOC 2012 and COCO benchmarks for WSSS.
• Demonstrated strong generalizability with promising results in weakly supervised object localization (WSOL) on the CUB-200-2011 dataset.

Conclusions:

• The proposed group-wise learning framework provides a foundation for more sophisticated and flexible semantic mining in WSSS.
• The method effectively leverages image-level annotations to generate reliable pseudo ground-truths for improved segmentation models.