Global-Guided Selective Context Network for Scene Parsing.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Semantic segmentation models often struggle with objects of varying scales.
• Existing methods apply multi-level context uniformly, neglecting pixel-specific context needs.

Purpose of the Study:

• To propose a novel network, the global-guided selective context network (GSCNet), for adaptive context selection in scene parsing.
• To improve semantic segmentation accuracy by addressing limitations in handling large and small objects.

Main Methods:

• Introduced two global-guided modules: Global-Guided Global Module (GGM) for selecting global context (GC) and Global-Guided Local Module (GLM) for local context (LC).
• Developed selective context blocks (SCBs) by combining GGM and GLM to propagate context hierarchically (coarse-to-fine).
• Integrated SCBs at different network levels to refine contextual information adaptively for each pixel.

Main Results:

• Achieved state-of-the-art performance on four benchmark datasets: Cityscapes, ADE20K, PASCAL Context, and COCO Stuff.
• GSCNet-101 attained 82.6% accuracy on the Cityscapes test set without using coarse data.
• GSCNet-101 achieved 56.22% accuracy on the ADE20K test set.

Conclusions:

• The proposed GSCNet effectively addresses the limitations of uniform context application in semantic segmentation.
• Adaptive context selection via GGM and GLM significantly enhances scene parsing accuracy across diverse datasets.
• GSCNet demonstrates superior performance, setting a new benchmark in semantic segmentation tasks.