HRNet: 3D Object Detection Network for Point Cloud with Hierarchical Refinement

Bin Lu*, Yang Sun, Zhenyu Yang, Ran Song, Haiyang Jiang, YONGHUAI LIU

*Corresponding author for this work

Research output: Contribution to journalArticle (journal)peer-review


Recently, 3D object detection from LiDAR point clouds has advanced rapidly. Although the second
stage can improve the detection performance significantly, prior works concern little about
the essential differences among different stages for the performance enhancement. To address
this, this paper proposes a Hierarchical Refinement Network (HRNet) with two novel strategies.
Firstly, we build the detection head on multi-scale voxel features to optimize the regression
branch progressively with an effective Scale-aware Attentive Propagation (SAP) module. Then,
we propose a Dynamic Sample Selection (DSS) module for the recalculation of the IoU during
each stage to obtain more balanced positive and negative sample selections. Experiments over
benchmark datasets show the effectiveness of our HRNet, particularly for car detection in the
sparse point clouds.
Original languageEnglish
Article number110254
Pages (from-to)1-11
JournalPattern Recognition
Issue number110254
Early online date5 Jan 2024
Publication statusPublished - 9 Jan 2024


  • 3D object detection
  • LiDAR point clouds
  • multi-scale features
  • transformer
  • dynamic sample selection
  • hierarchical refinement

Research Centres

  • Centre for Intelligent Visual Computing Research
  • Data and Complex Systems Research Centre


Dive into the research topics of 'HRNet: 3D Object Detection Network for Point Cloud with Hierarchical Refinement'. Together they form a unique fingerprint.

Cite this