Is Context-Aware CNN Ready for the Surroundings? Panoramic Semantic Segmentation in the Wild

Area of Science:

• Computer Vision
• Autonomous Transportation
• Machine Learning

Background:

• Semantic segmentation is crucial for autonomous driving, with Convolution Neural Networks (CNNs) achieving high accuracy using contextual information.
• Current CNNs are primarily evaluated on limited Field of View (FoV) pinhole images, not the increasingly popular panoramic imagery.

Purpose of the Study:

• To develop and evaluate semantic segmentation methods specifically for omnidirectional, wide-FoV panoramic images.
• To address the gap in comprehensive evaluation of semantic segmentation on panoramic data, which offers rich contextual information.

Main Methods:

• A novel concurrent horizontal and vertical attention module to exploit contextual priors in panoramic images.
• A multi-source omni-supervised learning scheme incorporating data distillation for panoramic domain adaptation.
• Introduction of the Wild PAnoramic Semantic Segmentation (WildPASS) dataset for evaluating CNNs on diverse, real-world panoramic scenes.

Main Results:

• The proposed attention module and learning scheme significantly improve accuracy for semantic segmentation on wide-FoV images.
• The WildPASS dataset facilitates robust evaluation, revealing performance gains over state-of-the-art methods in panoramic domains.
• The high-efficiency architecture achieved substantial accuracy improvements, demonstrating effectiveness in real-world navigation challenges.

Conclusions:

• The proposed methods effectively leverage panoramic contextual information for enhanced semantic segmentation in autonomous transportation.
• The developed dataset and learning scheme provide a strong foundation for future research in wide-FoV perception.
• This work advances the capabilities of semantic segmentation for real-world navigation applications using panoramic imagery.