Scholarlink

<p>In the 3D object detection, integrating depth information from point clouds with the semantic information from images is crucial for achieving high performance. Current methods often fall short in fully capturing the semantic nuances and intricate details of objects due to limited feature fusion. To address this challenge, we propose a global–local feature fusion 3D object detection network (GLFF3D). This network comprises two streams: a point cloud stream and an image stream. The image stream is used to extract global and local features from images. The point cloud stream includes two stages: Proposal-acquisition and Proposal-refinement. In the Proposal-acquisition stage, we fuse global image features with point cloud features to integrate semantic and spatial information comprehensively. In the Proposal-refinement stage, we fuse local image features with high-confidence object point features to enhance object localization and detail perception. We design a global feature multi-scale enhancement module (GFME) to strengthen the representation capability of global image features. We also design a global feature-aware fusion module global (GFAF) and a local feature-aware fusion module (LFAF) for effective feature integration. Experiment results on the SUN RGB-D dataset demonstrate that our model outperforms most existing 3D detection models, achieving a mean average precision (mAP) of 69.14% at an IoU threshold of 0.25 and 50.77% at an IoU threshold of 0.5, making a significant advan cement in 3D object detection. Code is available at <a href="https://github.com/zy-zhang66/GLFF3D">https://github.com/zy-zhang66/GLFF3D</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

Global–local feature fusion network for accurate 3D object detection in point clouds and images

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Haishun Du,
  • Zhengyang Zhang,
  • Wenzhe Zhang,
  • Linbing Cao

摘要

In the 3D object detection, integrating depth information from point clouds with the semantic information from images is crucial for achieving high performance. Current methods often fall short in fully capturing the semantic nuances and intricate details of objects due to limited feature fusion. To address this challenge, we propose a global–local feature fusion 3D object detection network (GLFF3D). This network comprises two streams: a point cloud stream and an image stream. The image stream is used to extract global and local features from images. The point cloud stream includes two stages: Proposal-acquisition and Proposal-refinement. In the Proposal-acquisition stage, we fuse global image features with point cloud features to integrate semantic and spatial information comprehensively. In the Proposal-refinement stage, we fuse local image features with high-confidence object point features to enhance object localization and detail perception. We design a global feature multi-scale enhancement module (GFME) to strengthen the representation capability of global image features. We also design a global feature-aware fusion module global (GFAF) and a local feature-aware fusion module (LFAF) for effective feature integration. Experiment results on the SUN RGB-D dataset demonstrate that our model outperforms most existing 3D detection models, achieving a mean average precision (mAP) of 69.14% at an IoU threshold of 0.25 and 50.77% at an IoU threshold of 0.5, making a significant advan cement in 3D object detection. Code is available at https://github.com/zy-zhang66/GLFF3D.