Scholarlink

<p>Multi-modal 3D object detection (3DOD) is fundamental to applications such as autonomous driving, robotics, and augmented reality, where robust scene understanding from heterogeneous sensor data is required. Despite substantial progress, existing 3DOD approaches often suffer from inefficient multi-modal fusion mechanisms and limited robustness when processing sparse and irregular point cloud observations. To address these challenges, we propose a unified framework that integrates a compact Gaussian representation based on Vector–Matrix decomposition (VMGS), a modality-aware feature extraction network (DualModNet), and a graph-based detection architecture (GraphQuery3D). The VMGS representation encodes 3D scenes in a structured and memory-efficient manner, reducing computational overhead while preserving geometric expressiveness. DualModNet extracts complementary modality-specific features from LiDAR and RGB inputs, facilitating structured cross-modal integration. Built upon the constructed Gaussian representation, GraphQuery3D performs detection via graph-based feature aggregation and query-driven localization. Extensive experiments demonstrate that the proposed framework consistently outperforms state-of-the-art baselines.</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

Multi-modal Gaussian splatting for 3D object detection

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Shuai Guo,
  • Qiyue Zhang,
  • Zhenpeng Yin,
  • Jiajun Lin,
  • Jingyi Wei,
  • Shangjing Sun,
  • Hongyang Bai

摘要

Multi-modal 3D object detection (3DOD) is fundamental to applications such as autonomous driving, robotics, and augmented reality, where robust scene understanding from heterogeneous sensor data is required. Despite substantial progress, existing 3DOD approaches often suffer from inefficient multi-modal fusion mechanisms and limited robustness when processing sparse and irregular point cloud observations. To address these challenges, we propose a unified framework that integrates a compact Gaussian representation based on Vector–Matrix decomposition (VMGS), a modality-aware feature extraction network (DualModNet), and a graph-based detection architecture (GraphQuery3D). The VMGS representation encodes 3D scenes in a structured and memory-efficient manner, reducing computational overhead while preserving geometric expressiveness. DualModNet extracts complementary modality-specific features from LiDAR and RGB inputs, facilitating structured cross-modal integration. Built upon the constructed Gaussian representation, GraphQuery3D performs detection via graph-based feature aggregation and query-driven localization. Extensive experiments demonstrate that the proposed framework consistently outperforms state-of-the-art baselines.