QDNet: Query-Denoising Network for Visual Traffic Knowledge Graph Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Transportation Systems

Background:

• Traffic scene perception is crucial for intelligent transportation systems but current methods lack comprehensive scene understanding.
• Existing approaches often focus on specific elements, failing to capture the holistic nature of traffic environments.

Purpose of the Study:

• To address the Visual Traffic Knowledge Graph Generation (VTKGG) task, aiming for a comprehensive representation of traffic scene information.
• To develop an end-to-end framework that integrates multiple subtasks for efficient knowledge graph generation.

Main Methods:

• Propose Query-Denoising Network (QDNet) to integrate subtasks via diverse queries, enabling end-to-end generation of visual traffic knowledge graphs.
• Incorporate a query-denoising strategy during training to enhance model robustness and performance by recovering ground truth from noisy inputs.

Main Results:

• QDNet effectively streamlines the generation of visual traffic knowledge graphs by eliminating intermediate steps.
• The query-denoising method significantly improves the accuracy, robustness, and overall performance of the model.
• Experiments demonstrate the superiority of the proposed framework over existing methods and its effectiveness on related tasks like Panoptic Scene Graph Generation.

Conclusions:

• The proposed QDNet framework offers a superior approach to Visual Traffic Knowledge Graph Generation.
• The query-denoising strategy is effective in enhancing the performance and robustness of multi-task learning models in computer vision.
• This work advances the comprehensive understanding and representation of traffic scenes for intelligent transportation systems.