Scholarlink

<p>This paper proposes a novel depth estimation model, ReDepthNet, based on the alignment of semantic segmentation masks for generating radar depth projections. The model aims to fuse image and radar data to achieve a high-precision depth estimation. A key innovation is the design of a Fusion Encoder that integrates the Efficient Multi-Scale Attention Module (EMA), which efficiently fuses RGB images and radar data. This approach not only captures global features, but also significantly improves the stripe-like artifacts in depth maps and enhances the clarity of object boundaries. Comprehensive experiments conducted on the nuScenes dataset show that, compared to our baseline model, ReDepthNet achieves MAE improvements of 8.2%, 5.5%, and 4.8% at distances of 50&#xa0;m, 70&#xa0;m, and 80&#xa0;m, respectively, and RMSE improvements of 8.2%, 5.4%, and 4.6%, showing significant performance advantages. Furthermore, our experiments highlight the importance of effectively utilizing semantic segmentation mask information and radar point distribution characteristics for depth estimation tasks.</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

ReDepthNet: a radar and camera depth estimation model based on semantic segmentation mask region alignment

  • 期刊论文

  • 发表日期：

    2025年8月5日

查看全文
  • Hong Liang,
  • Xu Zhang,
  • Qian Zhang,
  • Mingwen Shao

摘要

This paper proposes a novel depth estimation model, ReDepthNet, based on the alignment of semantic segmentation masks for generating radar depth projections. The model aims to fuse image and radar data to achieve a high-precision depth estimation. A key innovation is the design of a Fusion Encoder that integrates the Efficient Multi-Scale Attention Module (EMA), which efficiently fuses RGB images and radar data. This approach not only captures global features, but also significantly improves the stripe-like artifacts in depth maps and enhances the clarity of object boundaries. Comprehensive experiments conducted on the nuScenes dataset show that, compared to our baseline model, ReDepthNet achieves MAE improvements of 8.2%, 5.5%, and 4.8% at distances of 50 m, 70 m, and 80 m, respectively, and RMSE improvements of 8.2%, 5.4%, and 4.6%, showing significant performance advantages. Furthermore, our experiments highlight the importance of effectively utilizing semantic segmentation mask information and radar point distribution characteristics for depth estimation tasks.