Scholarlink

<p>Detecting three-dimensional (3D) objects from sparse LiDAR point clouds become difficult when trying to detect small and distant objects like pedestrians and cyclists. To address scale heterogeneity and geometric indirection in existing two-stage detectors, we propose multi-scale point transformers by region based convolutional neural network (MPTR-CNN), a novel framework that integrates multi-scale point transformers with region-based refinement. The region proposal network (RPN) in its stage-1 uses a multi-scale neighborhood embedding module together with a jump-connection offset attention mechanism to achieve dual functionality by capturing precise local details and intermediate structural elements and overall spatial information. The stage-2 region-based convolutional neural network (RCNN) refines proposals via a canonical six-feature fusion strategy that preserves absolute depth, reflectivity, and pose-invariant spatial cues. Extensive experiments at the Karlsruhe&#xa0;Institute of&#xa0;Technology and&#xa0;Toyota&#xa0;Technological&#xa0;Institute at Chicago (KITTI) and nuTonomy scenes (nuScenes) benchmarks demonstrate that MPTR-CNN achieves state-of-the-art detection accuracy (70.31% mAP on KITTI test set, 57.6% on nuScenes val.), with exceptional gains on pedestrians and cyclists. The model maintains a 49&#xa0;ms inference latency, establishing a new accuracy-speed Pareto frontier for real-time 3D perception.</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

3D object detection using a multi-scale point transformer-RCNN

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Malik Haris,
  • Guoqiang Zhang,
  • Liming Li,
  • Muhammad Shahid Mastoi,
  • Zhenqing Li,
  • Asif Raza

摘要

Detecting three-dimensional (3D) objects from sparse LiDAR point clouds become difficult when trying to detect small and distant objects like pedestrians and cyclists. To address scale heterogeneity and geometric indirection in existing two-stage detectors, we propose multi-scale point transformers by region based convolutional neural network (MPTR-CNN), a novel framework that integrates multi-scale point transformers with region-based refinement. The region proposal network (RPN) in its stage-1 uses a multi-scale neighborhood embedding module together with a jump-connection offset attention mechanism to achieve dual functionality by capturing precise local details and intermediate structural elements and overall spatial information. The stage-2 region-based convolutional neural network (RCNN) refines proposals via a canonical six-feature fusion strategy that preserves absolute depth, reflectivity, and pose-invariant spatial cues. Extensive experiments at the Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago (KITTI) and nuTonomy scenes (nuScenes) benchmarks demonstrate that MPTR-CNN achieves state-of-the-art detection accuracy (70.31% mAP on KITTI test set, 57.6% on nuScenes val.), with exceptional gains on pedestrians and cyclists. The model maintains a 49 ms inference latency, establishing a new accuracy-speed Pareto frontier for real-time 3D perception.