A Comprehensive Exploration of Fidelity Quantification in Computer-Generated Images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving

Background:

• Realistic road scene generation is vital for training and validating advanced driving systems.
• Existing methods for assessing synthetic data fidelity are often application-specific and lack a comprehensive framework.
• Computer-generated images (CGIs) present challenges in objective and subjective fidelity assessment.

Purpose of the Study:

• To propose a comprehensive conceptual framework for quantifying the fidelity of virtual RGB images.
• To develop a set of distinct metrics for assessing the realism of synthetic road scenes.
• To analyze statistical characteristics of real and synthetic road datasets for insights into perceived realism.

Main Methods:

• Analysis of local and global texture perspectives and high-frequency information in images.
• Statistical comparison of over 28,000 real and synthetic road scene images from multiple datasets.
• Evaluation from the perspective of an embedded camera, not the human eye.

Main Results:

• A novel set of objective metrics for quantifying image fidelity has been developed.
• Insights into texture patterns and high-frequency components that contribute to realism perception were revealed.
• Pioneering objective scores were applied to real, virtual, and improved virtual data.

Conclusions:

• The proposed metrics offer a comprehensive approach to quantifying CGI fidelity beyond application-specific needs.
• This work provides valuable insights for improving the realism of synthetic road scene datasets.
• The developed objective scores serve as a crucial asset for the scientific community in evaluating data realism.