Scholarlink

<p>This study presents a real-time deep learning framework for no-reference quality assessment of indoor 3D point clouds. The proposed method combines geometric descriptors-curvature, planarity, and sphericity-with LAB color features using a multimodal fusion strategy. Segmented point cloud patches are converted into pseudo-image tensors and fed into a fine-tuned VGG16 model to estimate local quality scores. These scores are aggregated into a Global Quality Index (GQI) to evaluate overall scene quality. The framework is implemented on a Raspberry Pi with a depth camera to enable on-device real-time processing. Unlike existing methods, this is one of the first frameworks to achieve real-time, no-reference point cloud quality assessment with edge device implementation. Experimental results validate its accuracy and robustness across various indoor conditions, demonstrating advantages over traditional and deep learning-based baselines.</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

Real-time quality assessment of indoor 2D and 3D point clouds using VGG16 with multimodal feature fusion

  • 期刊论文

  • 发表日期：

    2025年7月29日

查看全文
  • A. M. Deepthi Shree,
  • M. Evany Anne,
  • C. Sekar,
  • M. Brindha

摘要

This study presents a real-time deep learning framework for no-reference quality assessment of indoor 3D point clouds. The proposed method combines geometric descriptors-curvature, planarity, and sphericity-with LAB color features using a multimodal fusion strategy. Segmented point cloud patches are converted into pseudo-image tensors and fed into a fine-tuned VGG16 model to estimate local quality scores. These scores are aggregated into a Global Quality Index (GQI) to evaluate overall scene quality. The framework is implemented on a Raspberry Pi with a depth camera to enable on-device real-time processing. Unlike existing methods, this is one of the first frameworks to achieve real-time, no-reference point cloud quality assessment with edge device implementation. Experimental results validate its accuracy and robustness across various indoor conditions, demonstrating advantages over traditional and deep learning-based baselines.