Global-and-Local Context Network for Semantic Segmentation of Street View Images.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Semantic segmentation is crucial for autonomous vehicle scene understanding.
• Existing methods leverage Fully Convolutional Networks (FCNs) and local context but often neglect rich global information.
• A unified approach to integrate both global and local contextual information remains under-explored.

Purpose of the Study:

• To propose a novel network architecture, GLNet, for enhanced semantic segmentation of street view images.
• To systematically integrate global spatial information with dense local multi-scale context.
• To reduce segmentation errors by effectively modeling inter-object relationships within a scene.

Main Methods:

• Developed a global-and-local network architecture (GLNet).
• Incorporated global spatial information and dense local multi-scale context.
• Introduced a channel attention module for refining segmentation using low-level features.

Main Results:

• Achieved 80.8% test accuracy on the Cityscapes dataset.
• Demonstrated superior performance compared to existing state-of-the-art methods.
• GLNet effectively models object relationships, reducing segmentation errors.

Conclusions:

• GLNet offers a systematic approach to utilizing both global and local context for semantic segmentation.
• The proposed architecture significantly improves segmentation accuracy for autonomous driving applications.
• GLNet represents a notable advancement in scene understanding for intelligent vehicle systems.