Scholarlink

<p>Video salient object detection (VSOD) is crucial in computer vision, aiming to identify and segment the most salient objects in videos. Most existing methods rely solely on bimodal (RGB, optical flow) or tri-modal (RGB, optical flow, depth) information for salient object detection. Due to limited modal dimensions or insufficient exploitation of intermodal complementarity, they struggle to capture complete object features in complex scenes, resulting in constrained detection performance. This paper introduces a Tri-modal Complementary Fusion Network (TCFNet) for SOD in RGB-D videos, leveraging RGB, optical flow, and depth information. TCFNet comprises an attention-enhanced tri-modal feature interaction fusion module (ATFM) and a gate-enhanced feedback module (GFM). ATFM enhances feature interaction and fusion across modalities, while GFM aggregates and refines multi-level features. Experimental results on five authoritative datasets demonstrate TCFNet’s superior performance, significantly outperforming 18 existing VSOD methods, with a 5.1% gain in the <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(F_\beta ^{max}\)</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mi>F</mi> <mi>β</mi> <mrow> <mi mathvariant="italic">max</mi> </mrow> </msubsup> </math></EquationSource> </InlineEquation> metric on the VOS dataset. The code is published on <a href="https://github.com/yangchengbang/TCFNet.">https://github.com/yangchengbang/TCFNet.</a></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

Enhancing Salient Object Detection in RGB-D Videos via Tri-modal Complementary Fusion

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Chengbang Yang,
  • Anzhi Wang,
  • Chunhong Ren,
  • Yun Shao

摘要

Video salient object detection (VSOD) is crucial in computer vision, aiming to identify and segment the most salient objects in videos. Most existing methods rely solely on bimodal (RGB, optical flow) or tri-modal (RGB, optical flow, depth) information for salient object detection. Due to limited modal dimensions or insufficient exploitation of intermodal complementarity, they struggle to capture complete object features in complex scenes, resulting in constrained detection performance. This paper introduces a Tri-modal Complementary Fusion Network (TCFNet) for SOD in RGB-D videos, leveraging RGB, optical flow, and depth information. TCFNet comprises an attention-enhanced tri-modal feature interaction fusion module (ATFM) and a gate-enhanced feedback module (GFM). ATFM enhances feature interaction and fusion across modalities, while GFM aggregates and refines multi-level features. Experimental results on five authoritative datasets demonstrate TCFNet’s superior performance, significantly outperforming 18 existing VSOD methods, with a 5.1% gain in the \(F_\beta ^{max}\) F β max metric on the VOS dataset. The code is published on https://github.com/yangchengbang/TCFNet.