CGNet: A Light-Weight Context Guided Network for Semantic Segmentation.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Increasing demand for semantic segmentation on mobile devices.
• Current state-of-the-art models are too large for mobile deployment.
• Existing lightweight models neglect segmentation-specific characteristics.

Purpose of the Study:

• To propose a novel, lightweight, and efficient semantic segmentation network for mobile applications.
• To address the limitations of existing models in terms of parameter count and memory footprint.
• To enhance segmentation accuracy by effectively utilizing contextual information.

Main Methods:

• Development of a novel Context Guided (CG) block to efficiently learn joint local and surrounding features.
• Integration of the CG block into CGNet, capturing contextual information across all network stages.
• Design focused on reducing parameters and memory footprint for mobile suitability.

Main Results:

• CGNet significantly outperforms existing lightweight segmentation networks with an equivalent number of parameters.
• Achieved 64.8% mean IoU on the Cityscapes dataset with under 0.5 million parameters.
• Demonstrated effectiveness without post-processing or multi-scale testing on Cityscapes and CamVid datasets.

Conclusions:

• CGNet is an effective and efficient lightweight network for semantic segmentation on mobile devices.
• The proposed CG block successfully captures and utilizes contextual information for improved accuracy.
• CGNet offers a viable solution for deploying high-performance semantic segmentation in resource-constrained environments.