Scholarlink

<p>Salient Object Detection (SOD) is an important preprocessing operation in computer vision tasks. To pursue detection accuracy, most existing RGB-depth (RGB-D) SOD models often use bulky networks with large parameters and complex structures to predict salient objects. Few models simultaneously consider the issues of high efficiency and concise structure, which will greatly benefit computer vision tasks. To this end, we rethink the rationale of RGB-D SOD networks and argue that their performance mainly benefits from the global and local modeling capabilities obtained through the collaborative learning between the encoder and decoder. Specifically, we propose a <b>P</b>owerful and <b>E</b>fficient <b>Net</b>work (termed PENet) for RGB-D SOD. The philosophy behind PENet is a belief that great truths are always simple. The PENet adopts a pair of asymmetric backbone networks as encoders and a cleverly designed decoder consisting of Multi-modal Feature Fusion Modules (MFFM) and Multi-scale Feature Refinement Modules (MFRM). The PENet has 14.9M parameters, 16.0G Flops, and 43.4 FPS, but achieves state-of-the-art performance compared to 26 models on 7 datasets. Experimental results show that our PENet with a relatively simpler structure exhibits amazing performance.</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

Rethinking RGB-D salient object detection

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Chang Kou,
  • Jinyu Han,
  • Mengyin Wang

摘要

Salient Object Detection (SOD) is an important preprocessing operation in computer vision tasks. To pursue detection accuracy, most existing RGB-depth (RGB-D) SOD models often use bulky networks with large parameters and complex structures to predict salient objects. Few models simultaneously consider the issues of high efficiency and concise structure, which will greatly benefit computer vision tasks. To this end, we rethink the rationale of RGB-D SOD networks and argue that their performance mainly benefits from the global and local modeling capabilities obtained through the collaborative learning between the encoder and decoder. Specifically, we propose a Powerful and Efficient Network (termed PENet) for RGB-D SOD. The philosophy behind PENet is a belief that great truths are always simple. The PENet adopts a pair of asymmetric backbone networks as encoders and a cleverly designed decoder consisting of Multi-modal Feature Fusion Modules (MFFM) and Multi-scale Feature Refinement Modules (MFRM). The PENet has 14.9M parameters, 16.0G Flops, and 43.4 FPS, but achieves state-of-the-art performance compared to 26 models on 7 datasets. Experimental results show that our PENet with a relatively simpler structure exhibits amazing performance.