Scene Segmentation With Dual Relation-Aware Attention Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Scene segmentation is crucial for pixel-level recognition.
• Efficiently exploiting contextual information is essential for accurate scene segmentation.

Purpose of the Study:

• To propose a novel Dual Relation-aware Attention Network (DRANet) for enhanced scene segmentation.
• To adaptively capture contextual information using relation-aware attention mechanisms.

Main Methods:

• Developed a Dual Relation-aware Attention Network (DRANet) integrating spatial and channel attention modules.
• Employed self-attention mechanisms for adaptive context aggregation.
• Introduced compact attention modules to reduce computational cost.
• Incorporated a cross-level gating decoder to enhance spatial details.

Main Results:

• Achieved new state-of-the-art segmentation performance on Cityscapes, ADE20K, PASCAL Context, and COCO Stuff datasets.
• Obtained a Mean IoU score of 82.9% on the Cityscapes test set without using extra annotated data.
• Demonstrated the effectiveness of DRANet in adaptively capturing and exploiting contextual information.

Conclusions:

• DRANet effectively handles scene segmentation by adaptively capturing context.
• The proposed attention mechanisms and compact modules significantly improve performance.
• The network achieves superior results on challenging benchmarks, highlighting its practical applicability.