Scholarlink

<p>Learning-based multi-view depth estimation approaches have achieved remarkable success, primarily by accurately matching correspondences between the reference and source views to construct distinguishable cost volume. Existing methods using plane sweeping with ordered and preset sampling fail to make the initial cost volume discriminative. Additionally, insufficient cost aggregation exacerbates the challenges in achieving accurate depth estimation within low-texture regions and object boundaries. In this paper, we propose a novel multi-view depth estimation network, termed <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_3985_Article_IEq3.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="51" /> </InlineMediaObject> <EquationSource Format="TEX">\(\text {SP-A}\text {I}^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>SP-A</mtext> <msup> <mtext>I</mtext> <mn>2</mn> </msup> </mrow> </math></EquationSource> </InlineEquation> (sparse prior guided cost construction and adaptive intra and inter scale cost aggregation). Specifically, we propose a sparse prior guided strategy for dynamic and nonuniform sampling across a wide depth range to construct the cost volume, which introduces reasonable and fine-grained spatial partitioning to refine the depth with higher accuracy. Furthermore, to improve the depth quality under challenging regions, a novel adaptive intra and inter scale cost aggregation (<InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_3985_Article_IEq4.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="26" /> </InlineMediaObject> <EquationSource Format="TEX">\(\text {A}\text {I}^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>AI</mtext> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>-SA) module is proposed to enhance the power of feature representation. The proposed <InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_3985_Article_IEq5.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="51" /> </InlineMediaObject> <EquationSource Format="TEX">\({\hbox {SP-A}}\text {I}^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>SP-A</mtext> <msup> <mtext>I</mtext> <mn>2</mn> </msup> </mrow> </math></EquationSource> </InlineEquation> is trained end-to-end and experimental results demonstrate that our method achieves state-of-the-art results on various benchmark datasets.</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

SP-A\(\text {I}^{2}\): sparse prior guided cost construction and adaptive intra and inter scale cost aggregation for multi-view depth estimation

  • 期刊论文

  • 发表日期：

    2025年3月10日

查看全文
  • Qianqian Du,
  • Hui Yin,
  • Ming Han

摘要

Learning-based multi-view depth estimation approaches have achieved remarkable success, primarily by accurately matching correspondences between the reference and source views to construct distinguishable cost volume. Existing methods using plane sweeping with ordered and preset sampling fail to make the initial cost volume discriminative. Additionally, insufficient cost aggregation exacerbates the challenges in achieving accurate depth estimation within low-texture regions and object boundaries. In this paper, we propose a novel multi-view depth estimation network, termed \(\text {SP-A}\text {I}^{2}\) SP-A I 2 (sparse prior guided cost construction and adaptive intra and inter scale cost aggregation). Specifically, we propose a sparse prior guided strategy for dynamic and nonuniform sampling across a wide depth range to construct the cost volume, which introduces reasonable and fine-grained spatial partitioning to refine the depth with higher accuracy. Furthermore, to improve the depth quality under challenging regions, a novel adaptive intra and inter scale cost aggregation ( \(\text {A}\text {I}^{2}\) AI 2 -SA) module is proposed to enhance the power of feature representation. The proposed \({\hbox {SP-A}}\text {I}^{2}\) SP-A I 2 is trained end-to-end and experimental results demonstrate that our method achieves state-of-the-art results on various benchmark datasets.