Scholarlink

<p>No-reference point cloud quality assessment (PCQA) aims to predict perceptual quality without access to a pristine reference point cloud. Most existing PCQA methods typically formulate this task as a deterministic score regression problem, overlooking the fact that mean opinion scores (MOS) are obtained from a limited number of human subjective ratings and may therefore contain non-negligible subjective uncertainty. Inspired by these, we propose an uncertainty-aware multi-view post-aggregation framework, termed UM-PCQA, which predicts the point cloud quality score as the mean of the resulting posterior distribution. Firstly, a view-wise quality distribution prediction head estimates the mean and standard deviation of each projected view. Meanwhile, we utilize an attention module to adaptively estimate the importance weights of different views. Finally, the sample-level prediction is obtained by moment-matching of the view-wise means, variances, and view weights. Experimental results demonstrate that UM-PCQA achieves competitive performance against state-of-the-art methods on benchmarks and exhibits robustness in cross-dataset evaluation.</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

Uncertainty-aware multi-view post-aggregation for point cloud quality assessment

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Yonggang Xu,
  • Furong Pan,
  • Yangwei Li,
  • Xiaochuan Wang

摘要

No-reference point cloud quality assessment (PCQA) aims to predict perceptual quality without access to a pristine reference point cloud. Most existing PCQA methods typically formulate this task as a deterministic score regression problem, overlooking the fact that mean opinion scores (MOS) are obtained from a limited number of human subjective ratings and may therefore contain non-negligible subjective uncertainty. Inspired by these, we propose an uncertainty-aware multi-view post-aggregation framework, termed UM-PCQA, which predicts the point cloud quality score as the mean of the resulting posterior distribution. Firstly, a view-wise quality distribution prediction head estimates the mean and standard deviation of each projected view. Meanwhile, we utilize an attention module to adaptively estimate the importance weights of different views. Finally, the sample-level prediction is obtained by moment-matching of the view-wise means, variances, and view weights. Experimental results demonstrate that UM-PCQA achieves competitive performance against state-of-the-art methods on benchmarks and exhibits robustness in cross-dataset evaluation.