Scholarlink

<p>Depth completion is vital for various CV applications such as autonomous driving, robotics, and augmented reality. However, most of the existing approaches suffer from high computational overhead and slow inference speeds, limiting their real-time applicability. In this paper, we present the Linear Attention-based Cascade Hourglass Network (LA-CHN), a lightweight yet robust depth completion model designed for efficient and accurate dense depth prediction. The core of LA-CHN is our Lightweight Linear Attention (LLA) block, which substitutes quadratic self-attention with a ReLU-kernel linear mechanism and a spatial-reduction strategy to maintain a global receptive field at linear cost. These LLA blocks are embedded within a three-stage cascaded hourglass backbone, enabling multiscale feature aggregation and progressive refinement. Experimental results on the outdoor KITTI DC benchmark and indoor NYUv2 dataset show that our approach achieves superior performance compared to previous lightweight depth completion models.</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

Ultra fast-inference depth completion with linear attention-based cascaded hourglass network

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Zirui Wu,
  • Yongtao Hao

摘要

Depth completion is vital for various CV applications such as autonomous driving, robotics, and augmented reality. However, most of the existing approaches suffer from high computational overhead and slow inference speeds, limiting their real-time applicability. In this paper, we present the Linear Attention-based Cascade Hourglass Network (LA-CHN), a lightweight yet robust depth completion model designed for efficient and accurate dense depth prediction. The core of LA-CHN is our Lightweight Linear Attention (LLA) block, which substitutes quadratic self-attention with a ReLU-kernel linear mechanism and a spatial-reduction strategy to maintain a global receptive field at linear cost. These LLA blocks are embedded within a three-stage cascaded hourglass backbone, enabling multiscale feature aggregation and progressive refinement. Experimental results on the outdoor KITTI DC benchmark and indoor NYUv2 dataset show that our approach achieves superior performance compared to previous lightweight depth completion models.