{
  "paper_id": "7113f98c71be0ea4601f6cfbafb941e8",
  "paper_title": "CPGNet: Cascade Point-Grid Fusion Network for Real-Time LiDAR Semantic Segmentation",
  "paper_authors": [
    "Xiaoyan Li",
    "Gang Zhang",
    "Hongyu Pan",
    "Zhenhua Wang"
  ],
  "publication_date": "",
  "publication_venue": "unknown",
  "abstract": "LiDAR semantic segmentation essential for advanced autonomous driving is required to be accurate, fast, and easy-deployed on mobile platforms. Previous point-based or sparse voxel-based methods are far away from real-time applications since time-consuming neighbor searching or sparse 3D convolution are employed. Recent 2D projection-based methods, including range view and multi-view fusion, can run in real time, but suffer from lower accuracy due to information loss during the 2D projection. Besides, to improve the performance, previous methods usually adopt test time augmentation (TTA), which further slows down the inference process. To achieve a better speed-accuracy trade-off, we propose Cascade Point-Grid Fusion Network (CPGNet), which ensures both effectiveness and efficiency mainly by the following two techniques: 1) the novel Point-Grid (PG) fusion block extracts semantic features mainly on the 2D projected grid for efficiency, while summarizes both 2D and 3D features on 3D point for minimal information loss; 2) the proposed transformation consistency loss narrows the gap between the single-time model inference and TTA. The experiments on the SemanticKITTI and nuScenes benchmarks demonstrate that the CPGNet without ensemble models or TTA is comparable with the state-of-the-art RPVNet, while it runs 4.7 times faster.",
  "keywords": [
    "LiDAR语义分割",
    "实时推理",
    "点云",
    "鸟瞰图",
    "距离视图",
    "多视图融合",
    "自动驾驶",
    "变换一致性损失"
  ],
  "paper_institutions": [],
  "output_language": "zh",
  "provider": "doubao",
  "model": "glm-5.3",
  "prompt_template": "simple",
  "extracted_at": "2026-09-17T09:31:52.443838Z",
  "template_tag": "simple",
  "summary": "### 摘要\nLiDAR语义分割是高级自动驾驶的关键任务，要求模型同时具备准确性、实时性并易于部署到车载等移动平台。现有方法可分为三类：基于点的方法（PointNet、RandLA-Net等）直接处理原始点云无信息损失，但依赖耗时的最远点采样与k近邻搜索；基于稀疏体素的方法（Cylinder3D、RPVNet等）精度领先但计算昂贵、难以实时；基于2D投影的方法（RangeNet++、MPF等）速度快且易部署于TensorRT等推理框架，但2D投影信息损失导致精度偏低，且以往方法常依赖测试时增强（TTA）多次推理进一步拖慢速度。为此，本文提出级联点-网格融合网络CPGNet，其核心是Point-Grid（PG）融合块：P2G操作将3D点特征投影到鸟瞰图与距离视图网格（网格特征按max pooling聚合：G_{h,w,c}^{2D}=max_{k∈R_{h,w}} F_{k,c}^{3D}，并用CUDA atomicMax避免并行冲突），轻量2D FCN高效提取语义特征，G2P操作经双线性插值将网格特征传回3D点，Point Fusion模块以特征拼接加两层MLP融合点、鸟瞰图、距离视图三分支特征以保持完整点云信息；同时提出变换一致性损失L_tc=Σ_c|ŷ_c^raw−ŷ_c^aug|，缩小单次推理与TTA之间的性能差距。实验选用SemanticKITTI（43,552个扫描、19类）与nuScenes（1,000个场景、16类）两大权威基准以验证速度-精度折衷与跨数据集表现，模型在8块GPU上以batch size 16从头训练48轮约24小时，采用SGD及多种数据增强，评估在PyTorch FP32、NVIDIA RTX 2080Ti上进行。以mIoU为指标，CPGNet在不使用模型集成或TTA时于SemanticKITTI验证集取得65.9 mIoU、43 ms的最佳速度-精度折衷，测试集达68.3 mIoU，nuScenes验证集达76.9 mIoU，与当时最优RPVNet相当而速度快4.7倍，其中truck与traffic-sign类别大幅领先；消融实验验证了Point Fusion、点特征、双下采样块、注意力特征金字塔融合及变换一致性损失各自的有效性。未来可通过LiDAR与图像融合改善motorcyclist等少样本类别。实际部署方面，除P2G与G2P外所有组件可直接部署于TensorRT，二者可用高效CUDA实现，TensorRT FP16下每扫描仅26.8 ms且性能损失可忽略（-0.1 mIoU），满足车载移动平台的实时约束。\n\n### 关键要点\n- CPGNet以级联框架融合点、鸟瞰图与距离视图特征，结合基于点方法信息完整与2D投影方法速度快的优势，实现准确、快速、易部署的LiDAR语义分割。\n- PG融合块分四步：P2G将点特征投影到鸟瞰图与距离视图网格（max pooling聚合，CUDA atomicMax避免并行冲突），2D FCN提取语义特征，G2P用双线性插值将网格特征传回3D点，Point Fusion以特征拼接和两层MLP融合三分支特征。\n- 鸟瞰图忽略z维、距离视图丢弃r维，两视图互补以缓解2D投影信息损失；在SemanticKITTI上99.99%的点至少落入一个视图范围内。\n- 2D FCN采用仅9层、最大128通道的轻量ResNet骨干，占CPGNet超过90%的计算量；提出双下采样块（并行2D卷积与MaxPool）与注意力特征金字塔融合以自动选择不同层级特征。\n- 总损失为L_total=L_wce+2L_ls+L_tc：加权交叉熵以α_c=1/(F_c+ε)（ε=0.001）强调稀有类别，Lovász-Softmax直接优化mIoU，变换一致性损失约束原始与增强点云预测的一致性。\n- 变换一致性损失受TTA启发，训练时加入并在推理时移除TTA，比需四次推理的TTA略好（+0.2 mIoU）且仅需单次推理，显著节省推理时间。\n- 数据集：SemanticKITTI含43,552个360°扫描（训练19,130、验证4,071、在线测试20,351），采用19类评估；nuScenes含1,000个场景（训练28,130、验证6,019、测试6,008），采用16类评估。\n- 训练设置：8 GPU、batch size 16、从头训练48轮约24小时；SGD动量0.9、权重衰减0.001、初始学习率0.02每10轮衰减0.1；数据增强含z轴随机旋转、[0.95,1.05]全局缩放、x/y轴翻转、高斯噪声N(0,0.02)及instance CutMix。\n- 结果：SemanticKITTI验证集65.9 mIoU/43 ms（PyTorch FP32、RTX 2080Ti），测试集68.3 mIoU，nuScenes验证集76.9 mIoU，与最优RPVNet相当但快4.7倍，truck与traffic-sign类别大幅超越RPVNet。\n- 消融实验：Point Fusion较MPF后处理提升+0.9 mIoU，引入点特征提升+2.1 mIoU，两个PG融合块带来最大提升；双下采样块+0.3 mIoU、注意力特征金字塔融合+0.7 mIoU。\n- 部署约束：除P2G与G2P外组件可直接部署于TensorRT，P2G/G2P以高效CUDA实现；TensorRT FP16下每扫描26.8 ms，性能仅下降0.1 mIoU。\n- 局限：motorcyclist类别因训练样本少且与bicyclist、motorcycle混淆而表现较差，可通过LiDAR与图像融合解决。"
}