{
  "paper_id": "dee481e674a280afccbabdc7469c4f9e",
  "paper_title": "CluB: Cluster Meets BEV for LiDAR-Based 3D Object Detection",
  "paper_authors": [
    "Yingjie Wang",
    "Jiajun Deng",
    "Yuenan Hou",
    "Yao Li",
    "Yu Zhang",
    "Jianmin Ji",
    "Wanli Ouyang",
    "Yanyong Zhang"
  ],
  "publication_date": "",
  "publication_venue": "Advances in Neural Information Processing Systems",
  "abstract": "Currently, LiDAR-based 3D detectors are broadly categorized into two groups, namely, BEV-based detectors and cluster-based detectors. BEV-based detectors capture the contextual information from the Bird's Eye View (BEV) and fill their center voxels via feature diffusion with a stack of convolution layers, which, however, weakens the capability of presenting an object with the center point. On the other hand, cluster-based detectors exploit the voting mechanism and aggregate the foreground points into object-centric clusters for further prediction. In this paper, we explore how to effectively combine these two complementary representations into a unified framework. Specifically, we propose a new 3D object detection framework, referred to as CLuB, which incorporates an auxiliary cluster-based branch into the BEV-based detector by enriching the object representation at both feature and query levels. Technically, CLuB is comprised of two steps. First, we construct a cluster feature diffusion module to establish the association between cluster features and BEV features in a subtle and adaptive fashion. Based on that, an imitation loss is introduced to distill object-centric knowledge from the cluster features to the BEV features. Second, we design a cluster query generation module to leverage the voting centers directly from the cluster branch, thus enriching the diversity of object queries. Meanwhile, a direction loss is employed to encourage a more accurate voting center for each cluster. Extensive experiments are conducted on Waymo and nuScenes datasets, and our CLuB achieves state-of-the-art performance on both benchmarks.\n",
  "keywords": [
    "LiDAR-based 3D object detection",
    "BEV representation",
    "cluster-based representation",
    "feature fusion",
    "query expansion",
    "transformer",
    "Waymo Open Dataset",
    "nuScenes"
  ],
  "paper_institutions": [
    "University of Science and Technology of China",
    "Shanghai AI Laboratory",
    "The University of Adelaide, AIML"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "deepseek-v4-flash",
  "prompt_template": "simple",
  "extracted_at": "2026-08-18T10:30:45.053986Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n本文针对 LiDAR 三维目标检测中 BEV 表征与聚类表征互补但难以融合的问题展开研究。现有 BEV 检测器能捕获上下文信息，但物体中心点可能为空，需靠卷积扩散中心特征；聚类检测器通过投票机制聚合前景点形成物体中心簇，保留物体结构却依赖前景分割且缺乏上下文。为此，作者提出 CluB 框架，将辅助聚类分支集成到主流 BEV 检测器中，在特征层和查询层同时增强物体表征：特征层使用 Cluster Feature Diffusion (CFD) 模块，先将投票中心按公式 p_x=(c_i^x-B_x)/(v·d), p_y=(c_i^y-B_y)/(v·d) 映射到 BEV 平面得到 vote BEV，再通过类别感知扩散建立软性自适应关联，并引入 imitation loss 将物体中心知识从聚类特征蒸馏到 BEV 特征；查询层设计 Cluster Query Generation (CQG) 模块，从 vote BEV 激活图选取 top 聚类查询并与 BEV 查询拼接，同时用 direction loss 约束投票方向以提高投票中心精度。实验在 Waymo Open Dataset 和 nuScenes 上进行，采用相应点云范围、体素大小、数据增强与 12/36 epoch 训练设置；在 Waymo 验证集上 CluB 达到 73.4/71.4 mAP/mAPH (L2)，在 nuScenes 测试集上达到 71.2 NDS 和 66.0 mAP，均超越此前单帧方法，消融实验显示各组件共同带来 3.4% mAPH/L2 绝对提升。进一步探索方向是降低 BEV 特征图随感知范围二次增长的计算与空间复杂度，以支持长距离检测；该方法面向自动驾驶和机器人等真实场景，受限于 LiDAR 点云稀疏性、前景分割质量以及长距离感知开销。核心机制在于通过 CFD 的类别感知扩散和 CQG 的查询扩展，将上下文感知的 BEV 表征与物体中心的聚类表征统一到同一框架中。\n\n### 关键要点\n\n- CluB 将 BEV 检测器与辅助聚类分支结合，在特征层和查询层两个层面增强物体表征，以统一两种互补的 LiDAR 三维检测范式。\n- BEV 表征擅长捕获上下文信息，但物体中心点可能为空；聚类表征通过投票机制保留物体结构，但依赖前景分割且缺乏上下文。\n- CFD 模块先把聚类投票中心映射到 BEV 平面生成 vote BEV，再按预测类别用最大池化扩散到邻域，从而建立软性自适应的簇-BEV 特征关联。\n- 模仿损失对 vote BEV 与 dense BEV 施加掩码约束，将物体中心导向的聚类知识蒸馏到 BEV 分支，提升表征学习稳定性。\n- CQG 模块在 vote BEV 上生成激活图并选取 top 聚类查询，与 BEV 热力图得到的查询拼接，丰富 transformer 解码器的物体查询多样性。\n- 方向损失基于余弦相似度约束前景点到投票中心的偏移方向，使聚类查询对应的投票中心更准确。\n- 在 Waymo 验证集上，CluB 达到 73.4/71.4 mAP/mAPH (L2)，在 nuScenes 测试集上达到 71.2 NDS 和 66.0 mAP，均取得当时单帧 LiDAR 方法的最优结果。\n- 消融实验表明，CFD、模仿损失、CQG 和方向损失各自带来增益，完整 CluB 相比 BEV 基线提升 3.4% mAPH/L2。\n- 与直接拼接、2D 卷积扩散等策略相比，类别感知最大池化扩散在车辆和行人上分别提升 1.0% 和 1.4% APH/L2，验证了软性自适应关联的有效性。\n- 聚类模块设计上，KNN、ball query 和 SIR 等分组策略影响性能，SIR 实例级聚类取得最佳结果；CluB 的局限是 BEV 特征图复杂度随感知范围二次增长，长距离检测仍需高效化探索。"
}