Scholarlink

<p>Color-guided depth super-resolution (DSR) addresses the challenging task of reconstructing high-resolution (HR) depth maps from their low-resolution (LR) counterparts using aligned HR color images as guidance. While color images provide valuable structural cues, direct cross-modal feature transfer often yields unsatisfactory results due to fundamental discrepancies in structural representations between depth and color modalities. To overcome this limitation, we present a novel Structure Selection and Modulation Network (SSMNet) that strategically leverages color guidance through two key innovations: (1) a Structural Representation Selection (SRS) module employing a pretrained structure encoder with geometric constraints to precisely extract depth-relevant structural features, and (2) a Structural Representation Modulation (SRM) module that dynamically integrates these features at multiple super-resolution stages. Our extensive evaluations across both synthetic and real-world benchmarks demonstrate that SSMNet consistently outperforms existing state-of-the-art DSR methods, achieving superior reconstruction quality while effectively mitigating texture transfer artifacts.</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

Color-guided depth super-resolution with structure selection and modulation

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Xin Sun,
  • Xinchen Ye,
  • Haojie Li

摘要

Color-guided depth super-resolution (DSR) addresses the challenging task of reconstructing high-resolution (HR) depth maps from their low-resolution (LR) counterparts using aligned HR color images as guidance. While color images provide valuable structural cues, direct cross-modal feature transfer often yields unsatisfactory results due to fundamental discrepancies in structural representations between depth and color modalities. To overcome this limitation, we present a novel Structure Selection and Modulation Network (SSMNet) that strategically leverages color guidance through two key innovations: (1) a Structural Representation Selection (SRS) module employing a pretrained structure encoder with geometric constraints to precisely extract depth-relevant structural features, and (2) a Structural Representation Modulation (SRM) module that dynamically integrates these features at multiple super-resolution stages. Our extensive evaluations across both synthetic and real-world benchmarks demonstrate that SSMNet consistently outperforms existing state-of-the-art DSR methods, achieving superior reconstruction quality while effectively mitigating texture transfer artifacts.