Semantic Segmentation and Depth Estimation Based on Residual Attention Mechanism.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Semantic segmentation and depth estimation are vital for autonomous driving scene understanding.
• Jointly learning these tasks improves overall scenario comprehension.
• Existing methods using task-specific networks with shared backbones show limitations in global feature extraction.

Purpose of the Study:

• To propose a novel Multi-Task Residual Attention Network (MTRAN) for joint semantic segmentation and depth estimation.
• To enhance the extraction of global features for improved scene understanding in autonomous driving.
• To address the inadequacy of task-specific networks in multi-task learning scenarios.

Main Methods:

• Developed a Multi-Task Residual Attention Network (MTRAN) with a shared global network and dedicated attention networks for each task.
• Incorporated a Convolutional Block Attention Module (CBAM) to refine global feature maps.
• Utilized residual connections to mitigate network degradation.
• Implemented a random-weighted strategy within an impartial multi-task learning framework to balance task contributions.

Main Results:

• The proposed MTRAN effectively improves performance in both semantic segmentation and depth estimation tasks.
• Attention mechanisms enhance the focus on relevant global features.
• The random-weighted strategy ensures balanced learning across tasks, preventing dominance by a single task.
• Experimental results validate the effectiveness of the MTRAN architecture and learning strategy.

Conclusions:

• The MTRAN architecture offers a superior approach for joint semantic segmentation and depth estimation in autonomous driving.
• Integrating attention modules and a balanced multi-task learning strategy leads to more robust scene understanding.
• This work provides a significant advancement in developing more capable autonomous driving systems.