Scholarlink

<p>This paper presents a novel approach for predicting the simulation-to-real-world (sim2real) transferability of synthetic-trained vision models by analyzing foreground image quality. Image quality assessment (IQA) techniques are applied to paired synthetic and real-world image subsets to quantify visual similarity and identify feature-level limitations that correlate with transfer performance. Among the evaluated metrics, the Structural Similarity (SSIM) index and Complex Wavelet SSIM (CW-SSIM) index showed consistent trends, where higher IQA scores corresponded to improved sim2real detection accuracy–most notably with increases of approximately 0.10 (SSIM) and 0.15 (CW-SSIM). These metrics effectively captured contrast, structure, and luminance similarities, offering a practical proxy for assessing digital twin fidelity. To support reproducibility and broader use, we introduce the Synthetic Image Quality Analysis Calculator (SIQAC), an open-source tool for automated IQA evaluation and sim2real potential prediction across classifiers and object detectors. Additional experiments demonstrated that the IQA-based approach generalizes to real2sim scenarios using zero-shot object detectors. This work bridges concepts from the human visual system and compression analysis to provide a lightweight, interpretable method for early-stage validation of virtual autonomy pipelines.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting Transfer Performance Using Foreground Image Quality in Synthetic Vision Systems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月26日

查看全文
  • Michael A. Mardikes,
  • John T. Evans,
  • Nathan C. Sprague,
  • Gregory M. Shaver

摘要

This paper presents a novel approach for predicting the simulation-to-real-world (sim2real) transferability of synthetic-trained vision models by analyzing foreground image quality. Image quality assessment (IQA) techniques are applied to paired synthetic and real-world image subsets to quantify visual similarity and identify feature-level limitations that correlate with transfer performance. Among the evaluated metrics, the Structural Similarity (SSIM) index and Complex Wavelet SSIM (CW-SSIM) index showed consistent trends, where higher IQA scores corresponded to improved sim2real detection accuracy–most notably with increases of approximately 0.10 (SSIM) and 0.15 (CW-SSIM). These metrics effectively captured contrast, structure, and luminance similarities, offering a practical proxy for assessing digital twin fidelity. To support reproducibility and broader use, we introduce the Synthetic Image Quality Analysis Calculator (SIQAC), an open-source tool for automated IQA evaluation and sim2real potential prediction across classifiers and object detectors. Additional experiments demonstrated that the IQA-based approach generalizes to real2sim scenarios using zero-shot object detectors. This work bridges concepts from the human visual system and compression analysis to provide a lightweight, interpretable method for early-stage validation of virtual autonomy pipelines.