Scholarlink

<p>While voxel-based 3D detectors utilizing sparse convolution have achieved significant milestones in point cloud object detection, their modeling of long-range spatial topologies remains constrained by the inherent local receptive fields of convolutional kernels. Furthermore, existing shallow fusion strategies fail to adequately model complex cross-modal interactions, thereby hindering the exhaustive exploitation of complementary semantic information from 2D images. To address these challenges, we propose MIHP3D, a 3D object detection framework integrating multi-level interaction and hierarchical perception. Specifically, the framework introduces the Hierarchical Grouped Linear Encoding (HGLE) and Multi-Level Feature Aggregation (MLFA) modules, which transcend the limitations of local computation by establishing robust multi-scale representations and long-range dependencies, thus enhancing the holistic scene understanding and perceptual sensitivity. Building upon this, for multimodal scenarios, a Context-Aware Cross-Feature Fusion (CACF) module is integrated to facilitate deep synergy between high-level image semantics and local geometric structures. Extensive experiments on the SUN RGB-D, ScanNetV2, and S3DIS benchmarks demonstrate that MIHP3D significantly outperforms baseline models. Notably, the model achieves competitive results on the SUN RGB-D and S3DIS datasets, further validating its effectiveness.</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

MIHP3D: Multi-level interaction and hierarchical perception for 3D object detection

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Chengyi Zhang,
  • Shuang Wang,
  • Shengjie Ren,
  • Linxian Zhu,
  • Miaohui Zhang

摘要

While voxel-based 3D detectors utilizing sparse convolution have achieved significant milestones in point cloud object detection, their modeling of long-range spatial topologies remains constrained by the inherent local receptive fields of convolutional kernels. Furthermore, existing shallow fusion strategies fail to adequately model complex cross-modal interactions, thereby hindering the exhaustive exploitation of complementary semantic information from 2D images. To address these challenges, we propose MIHP3D, a 3D object detection framework integrating multi-level interaction and hierarchical perception. Specifically, the framework introduces the Hierarchical Grouped Linear Encoding (HGLE) and Multi-Level Feature Aggregation (MLFA) modules, which transcend the limitations of local computation by establishing robust multi-scale representations and long-range dependencies, thus enhancing the holistic scene understanding and perceptual sensitivity. Building upon this, for multimodal scenarios, a Context-Aware Cross-Feature Fusion (CACF) module is integrated to facilitate deep synergy between high-level image semantics and local geometric structures. Extensive experiments on the SUN RGB-D, ScanNetV2, and S3DIS benchmarks demonstrate that MIHP3D significantly outperforms baseline models. Notably, the model achieves competitive results on the SUN RGB-D and S3DIS datasets, further validating its effectiveness.