Scholarlink

Virtual Reality (VR) relies heavily on perceived realism to engage users. However, traditional assessment methods like questionnaires are time-consuming and expensive. This study investigates Deep Neural Networks (DNNs), specifically Convolutional Neural Networks (CNNs), to automatically assess VR scene realism from single images. We built a dataset of VR scenes with varying realism levels achieved by manipulating object texture and lighting resolution. We trained CNNs to extract features correlated to perceived realism and predict the corresponding scores. Using ResNet-18 as the backbone network, we achieved a strong performance with a Pearson’s linear correlation coefficient of 0.8029. This study demonstrates the promise of DNNs for efficient and objective VR scene realism assessment, potentially revolutionizing VR development and leading to more immersive experiences.

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

Investigating the Use of Deep Neural Networks for Predicting Perceived Realism in VR Scenes

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Peerawat Pannattee,
  • Shogo Shimada,
  • Vibol Yem,
  • Nobuyuki Nishiuchi

摘要

Virtual Reality (VR) relies heavily on perceived realism to engage users. However, traditional assessment methods like questionnaires are time-consuming and expensive. This study investigates Deep Neural Networks (DNNs), specifically Convolutional Neural Networks (CNNs), to automatically assess VR scene realism from single images. We built a dataset of VR scenes with varying realism levels achieved by manipulating object texture and lighting resolution. We trained CNNs to extract features correlated to perceived realism and predict the corresponding scores. Using ResNet-18 as the backbone network, we achieved a strong performance with a Pearson’s linear correlation coefficient of 0.8029. This study demonstrates the promise of DNNs for efficient and objective VR scene realism assessment, potentially revolutionizing VR development and leading to more immersive experiences.