BASeg: Boundary aware semantic segmentation for autonomous driving.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Deep Learning

Background:

• Semantic segmentation is crucial for autonomous driving, enabling street scene understanding.
• Existing methods often struggle with long-range dependencies between object interiors and boundaries.
• Refining object details and ensuring intra-class consistency remain challenges.

Purpose of the Study:

• To introduce a novel Boundary Aware Network (BASeg) for enhanced semantic segmentation.
• To exploit boundary information as a key cue for guiding context aggregation.
• To improve the accuracy and consistency of semantic segmentation in autonomous driving scenarios.

Main Methods:

• Developed a Boundary Aware Network (BASeg) incorporating a Boundary Refined Module (BRM) and Context Aggregation Module (CAM).
• BRM refines low-level boundary features using high-level semantic features.
• CAM captures long-range dependencies between boundary regions and inner object pixels.

Main Results:

• Achieved 45.72% mIoU on ADE20K, 81.2% on Cityscapes, and 77.3% on CamVid.
• Demonstrated superior performance compared to state-of-the-art ResNet101-based methods.
• The proposed method shows effectiveness and can be integrated with various CNN backbones.

Conclusions:

• BASeg effectively utilizes boundary information to enhance semantic segmentation accuracy.
• The network improves intra-class consistency and captures long-range object dependencies.
• The approach offers a computationally efficient method for improving autonomous driving perception.