Event-Based Semantic Segmentation With Posterior Attention.

Area of Science:

• Computer Vision
• Robotics
• Sensor Technology

Background:

• Transformers have advanced semantic segmentation, but performance degrades in poor lighting.
• Frame-based cameras have limited frame rates, unsuitable for real-time autonomous driving.
• Event cameras provide high-speed, high dynamic range data, ideal for challenging conditions, but lack mature algorithms.

Purpose of the Study:

• To develop advanced algorithms for event-based semantic segmentation.
• To leverage unique event camera data characteristics for improved perception.
• To enhance semantic segmentation in low-light and high-speed scenarios for autonomous systems.

Main Methods:

• Proposed a novel posterior attention module integrating event data priors with standard attention mechanisms.
• Developed EvSegFormer, an event-based semantic segmentation network by adapting the SegFormer architecture.
• Evaluated the model on MVSEC and DDD-17 datasets for event-based segmentation.

Main Results:

• EvSegFormer achieved state-of-the-art performance on event-based semantic segmentation tasks.
• The posterior attention module effectively utilizes event data characteristics, particularly for moving objects.
• Demonstrated the potential of event cameras for robust perception in challenging environments.

Conclusions:

• Event cameras and novel algorithms like EvSegFormer can overcome limitations of traditional cameras in autonomous driving.
• The posterior attention module offers a flexible approach to enhance existing segmentation backbones for event data.
• Further research into event-based vision algorithms is crucial for advancing autonomous systems.