Scholarlink

<p>Multimodal fusion from cameras and LiDAR is essential for 3D object detection in autonomous vehicles. Most methods separately extract image and point cloud features, generate sparse proposals, and perform cross-modal fusion at the candidate level. However, they overlook spatial long-range dependencies and inter-channel dependencies during image feature extraction. Moreover, when leveraging Transformer decoders for fusion, they often fail to mitigate over-smoothing and representation collapse. To address these issues, we propose PCFusion, a multimodal 3D object detection framework incorporating our proposed ParScaleNet and ContraSpaceOpt modules. ParScaleNet employs a hierarchical feature pyramid architecture combined with 1D convolutional localization for efficient multi-scale feature extraction. ContraSpaceOpt integrates implicit uniformity regularization from contrastive learning to enforce a uniform feature distribution in the embedding space, effectively preventing feature collapse. Experimental results on the nuScenes test set show that PCFusion achieves 72.3 mAP and 74.5 NDS, demonstrating consistent performance gains.</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

PCFusion: a parallel multi-scale feature extraction and contrastive optimization framework for camera–LiDAR 3D object detection

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Dongxuan Xie,
  • Youkang Zhang,
  • Shilong Xu,
  • Shufan Bi,
  • Nan Wang,
  • Xiangkun He

摘要

Multimodal fusion from cameras and LiDAR is essential for 3D object detection in autonomous vehicles. Most methods separately extract image and point cloud features, generate sparse proposals, and perform cross-modal fusion at the candidate level. However, they overlook spatial long-range dependencies and inter-channel dependencies during image feature extraction. Moreover, when leveraging Transformer decoders for fusion, they often fail to mitigate over-smoothing and representation collapse. To address these issues, we propose PCFusion, a multimodal 3D object detection framework incorporating our proposed ParScaleNet and ContraSpaceOpt modules. ParScaleNet employs a hierarchical feature pyramid architecture combined with 1D convolutional localization for efficient multi-scale feature extraction. ContraSpaceOpt integrates implicit uniformity regularization from contrastive learning to enforce a uniform feature distribution in the embedding space, effectively preventing feature collapse. Experimental results on the nuScenes test set show that PCFusion achieves 72.3 mAP and 74.5 NDS, demonstrating consistent performance gains.