Scholarlink

<p>Small-object detection remains challenging in LiDAR-based 3D perception due to sparse point observations, blurred object boundaries, and interference from surrounding background or nearby objects. In center-based, anchor-free detectors, these challenges become particularly evident in the second-stage refinement process, where fixed feature sampling patterns may fail to provide reliable object-specific evidence for small and distant targets. To address this problem, we propose Dynamic Offset Sampling (DOS), a refinement-oriented sampling mechanism that dynamically predicts multiple object-centric offsets around detected centers. Instead of extracting features from predefined locations, DOS adaptively selects feature positions according to local BEV context, providing more reliable refinement evidence for sparse objects. To further preserve geometric semantics after dynamic sampling, we introduce an Orientation-aware Offset Positional Embedding (OOPE), which jointly encodes Cartesian offsets and polar-coordinate cues to provide translation- and orientation-aware spatial information for box refinement. In addition, a lightweight BEV feature enhancement module and an offset boundary regularization term are incorporated to improve feature discrimination and stabilize offset prediction during training. Experiments on the nuScenes benchmark show that the proposed method improves small-object detection and orientation estimation over the CenterPoint baseline while maintaining comparable inference efficiency. Additional analyses, including KITTI cross-dataset validation, cross-architecture experiments, ablation studies, complexity evaluation, offset distribution analysis, and qualitative visualization, further support the effectiveness of the proposed refinement strategy.</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

Dynamic offset sampling for small object detection in LiDAR–based 3D perception

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月15日

查看全文
  • Jinjin Gao,
  • Jinlong Yan,
  • Ligang He,
  • Xiaoning Sun,
  • Huiyan Han

摘要

Small-object detection remains challenging in LiDAR-based 3D perception due to sparse point observations, blurred object boundaries, and interference from surrounding background or nearby objects. In center-based, anchor-free detectors, these challenges become particularly evident in the second-stage refinement process, where fixed feature sampling patterns may fail to provide reliable object-specific evidence for small and distant targets. To address this problem, we propose Dynamic Offset Sampling (DOS), a refinement-oriented sampling mechanism that dynamically predicts multiple object-centric offsets around detected centers. Instead of extracting features from predefined locations, DOS adaptively selects feature positions according to local BEV context, providing more reliable refinement evidence for sparse objects. To further preserve geometric semantics after dynamic sampling, we introduce an Orientation-aware Offset Positional Embedding (OOPE), which jointly encodes Cartesian offsets and polar-coordinate cues to provide translation- and orientation-aware spatial information for box refinement. In addition, a lightweight BEV feature enhancement module and an offset boundary regularization term are incorporated to improve feature discrimination and stabilize offset prediction during training. Experiments on the nuScenes benchmark show that the proposed method improves small-object detection and orientation estimation over the CenterPoint baseline while maintaining comparable inference efficiency. Additional analyses, including KITTI cross-dataset validation, cross-architecture experiments, ablation studies, complexity evaluation, offset distribution analysis, and qualitative visualization, further support the effectiveness of the proposed refinement strategy.