Coupling Global Context and Local Contents for Weakly-Supervised Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly-supervised semantic segmentation (WSSS) offers advantages in annotation efficiency and performance.
• Single-stage WSSS (SS-WSSS) emerged to address computational costs and complexity of multi-stage methods.
• Existing SS-WSSS models struggle with incomplete object and background segmentation due to insufficient context.

Purpose of the Study:

• To propose a novel single-stage WSSS model, the weakly supervised feature coupling network (WS-FCN).
• To address background and object incompleteness in SS-WSSS by capturing global context and local details.
• To achieve state-of-the-art performance using only image-level class label supervision.

Main Methods:

• Developed WS-FCN, a single-stage WSSS model utilizing image-level labels.
• Introduced a Flexible Context Aggregation (FCA) module to capture global object context across granular spaces.
• Proposed a Semantically Consistent Feature Fusion (SF2) module for bottom-up, parameter-learnable aggregation of local features.
• Employed a self-supervised, end-to-end training strategy.

Main Results:

• WS-FCN achieved state-of-the-art mean Intersection over Union (mIoU) scores.
• Achieved 65.02% mIoU on PASCAL VOC 2012 validation set and 64.22% on the test set.
• Obtained 34.12% mIoU on the MS COCO 2014 validation set.
• Demonstrated effectiveness and efficiency in experimental results.

Conclusions:

• The proposed WS-FCN effectively addresses limitations of previous SS-WSSS models.
• The FCA and SF2 modules successfully capture multiscale context and fine-grained local information.
• WS-FCN offers a promising direction for efficient and accurate weakly-supervised semantic segmentation.