Scholarlink

<p>To understand the semantic segmentation task of complex indoor scenes more accurately, due to the information difference between RGB and depth modalities and the difficulty of effectively fusing RGB and depth image information, and to address the above problems, we propose a new RGB-D cross-modal feature semantic segmentation network (PCBNet) based on positional crossing and broad features. The network firstly extracts the positional correlation information between different modalities adaptively by the positional cross-attention regulation module (PCARM), so as to effectively align the RGB modalities with their complementary modalities. Secondly, the broad feature fusion module (BFFM) is proposed to further enhance and fuse the aligned features to mine long-distance contextual information to capture and integrate cross-modal feature information more comprehensively. Finally, the final segmentation prediction results are obtained by sending both low-level and high-level features together to the decoder for multi-scale information fusion. Test results on two public datasets, NYU Depth V2 (NYUDV2) and SUN RGB-D, show that our network has a good segmentation performance. Among them, the mean intersection and merger ratio (mIoU) can reach 52.6%&#xa0;55.0% and 47.5%&#xa0;50.1%, respectively.</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

PCBNet: positional crossing and broad features network for indoor scene semantic segmentation

  • 期刊论文

  • 发表日期：

    2025年10月21日

查看全文
  • Huifang Hou,
  • Zhang Wenwen,
  • Zihao Zhang,
  • Sun Wentao,
  • Yale Yang,
  • Liu Gang,
  • Han Haipeng,
  • Liu Tianyuan,
  • Li Xiaofeng

摘要

To understand the semantic segmentation task of complex indoor scenes more accurately, due to the information difference between RGB and depth modalities and the difficulty of effectively fusing RGB and depth image information, and to address the above problems, we propose a new RGB-D cross-modal feature semantic segmentation network (PCBNet) based on positional crossing and broad features. The network firstly extracts the positional correlation information between different modalities adaptively by the positional cross-attention regulation module (PCARM), so as to effectively align the RGB modalities with their complementary modalities. Secondly, the broad feature fusion module (BFFM) is proposed to further enhance and fuse the aligned features to mine long-distance contextual information to capture and integrate cross-modal feature information more comprehensively. Finally, the final segmentation prediction results are obtained by sending both low-level and high-level features together to the decoder for multi-scale information fusion. Test results on two public datasets, NYU Depth V2 (NYUDV2) and SUN RGB-D, show that our network has a good segmentation performance. Among them, the mean intersection and merger ratio (mIoU) can reach 52.6% 55.0% and 47.5% 50.1%, respectively.