Scholarlink

<p>The existing stereoscopic methods have shown proficiency in reconstructing objects such as human faces and animals. However, they frequently encounter difficulties in accurately matching complex highlight regions due to information loss. To surmount this challenge, this paper introduces an innovative disparity correction strategy that leverages detection mask technology to effectively mitigate highlight effects. Our method encompasses an end-to-end network consisting of two principal modules: highlight detection and highlight removal. In the highlight detection stage, our approach generates a segmentation mask for highlight regions by analyzing both the input image and the designated region of interest. Following this, the highlight removal stage involves the integration of the original image with the highlight partition mask, utilizing partial convolution to produce an image with reduced highlight effects. Post the stereo matching process, the modified image is compared with the original, illustrating enhanced performance in stereo matching task. To support robust model training and evaluation, we have developed a specialized binocular vision stereo matching dataset that includes nine distinct highlight scenarios. Experimental results corroborate the effectiveness of our proposed method in addressing the challenges associated with highlight-induced matching discrepancies.</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

3D Surface Highlight Removal Method Based on Detection Mask

  • 期刊论文

  • 发表日期：

    2025年8月28日

查看全文
  • Jing Chen,
  • Leyun Sun,
  • Ying Song,
  • Yongning Geng,
  • Haitao Xu,
  • Wenqiang Xu

摘要

The existing stereoscopic methods have shown proficiency in reconstructing objects such as human faces and animals. However, they frequently encounter difficulties in accurately matching complex highlight regions due to information loss. To surmount this challenge, this paper introduces an innovative disparity correction strategy that leverages detection mask technology to effectively mitigate highlight effects. Our method encompasses an end-to-end network consisting of two principal modules: highlight detection and highlight removal. In the highlight detection stage, our approach generates a segmentation mask for highlight regions by analyzing both the input image and the designated region of interest. Following this, the highlight removal stage involves the integration of the original image with the highlight partition mask, utilizing partial convolution to produce an image with reduced highlight effects. Post the stereo matching process, the modified image is compared with the original, illustrating enhanced performance in stereo matching task. To support robust model training and evaluation, we have developed a specialized binocular vision stereo matching dataset that includes nine distinct highlight scenarios. Experimental results corroborate the effectiveness of our proposed method in addressing the challenges associated with highlight-induced matching discrepancies.