{
  "paper_id": "d2f526e35a985c58d0c48aea9f41cfe3",
  "paper_title": "LGD-BEV: Label-Guided Distillation for BEV 3D Object Detection",
  "paper_authors": [
    "Ji-Yong Lee",
    "Minho Park",
    "Youngjoo Jo",
    "Dong-Oh Kang"
  ],
  "publication_date": "2025",
  "publication_venue": "IEEE Access",
  "abstract": "Knowledge distillation in autonomous driving typically involves transferring information from high-cost sensor modalities such as LiDAR to camera-based models to improve 3D object detection. However, these methods often require access to additional modalities during training or inference, leading to increased engineering complexity and extensive trial-and-error efforts. This limitation hinders their scalability and deployment in real-world applications. In contrast, our approach leverages only 3D ground-truth annotations as an auxiliary intermediate supervision, acting as a lightweight teacher, thereby removing the dependency on external sensors or strong pre-trained teachers. In this paper, we introduce a Label Encoder module to generate dense semantic representations in the Bird's-Eye View (BEV) space to guide the training of camera-based networks through knowledge distillation. Specifically, we train an AutoEncoder on 3D ground-truth annotations and use its encoder to produce informative BEV features, which are then used as an auxiliary loss for the camera-based detection backbone. This yields a lightweight, modality-independent, and scalable solution, as demonstrated by extensive evaluations on the NuScenes dataset.\n",
  "keywords": [
    "AutoEncoder",
    "autonomous driving",
    "knowledge distillation",
    "multi-view cameras",
    "object detection"
  ],
  "paper_institutions": [
    "Electronics and Telecommunications Research Institute (ETRI)",
    "The University of Texas at Austin"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "deepseek-v4-flash",
  "prompt_template": "simple",
  "extracted_at": "2026-08-18T10:30:18.267626Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n本文针对自动驾驶中相机-only 3D目标检测性能落后于LiDAR方法的问题，提出LGD-BEV标签引导蒸馏框架。与依赖LiDAR等额外传感器或强预训练教师模型的传统跨模态知识蒸馏不同，该方法仅使用3D ground-truth标注训练一个轻量级Label Encoder（自编码器），在BEV空间中生成稠密语义表示，并通过特征级蒸馏损失监督相机检测骨干网络。核心机制可表示为 L_feat = (1/N_p) * sum M_ij || α(F_img_ij) - F_label_ij||_2^2，其中M为前景掩码，α为通道对齐模块；同时引入Gating Fusion Module动态融合类别语义与几何框特征，并采用Class-Aware Rendering按类别频率优先级避免稀有类别被覆盖。实验在nuScenes数据集上进行，基于BEVDepth和ResNet-50，输入分辨率256×704，训练24轮；在验证集上达到40.2 mAP和51.5 NDS，超过或接近使用CBGS和LiDAR蒸馏的复杂方法，且推理时无额外开销。消融实验表明GFM和CaR均带来提升，特征通道划分中图像特征占比越高效果越好，近处和远处目标均有改善。未来工作可探索半监督或合成标签场景，并需注意方法效果依赖标注质量。该方法适用于无额外传感器、低成本可扩展的真实自动驾驶部署场景。\n\n### 关键要点\n\n- LGD-BEV提出仅利用3D ground-truth标注作为轻量教师，不依赖LiDAR等外部传感器或预训练教师模型，从而降低训练复杂度和部署成本。\n- Label Encoder采用自编码器训练，将每个目标的类别和3D框属性分别经MLP嵌入，并通过可学习的门控融合生成BEV标签特征。\n- Gating Fusion Module通过sigmoid门控权重γ对类别特征和几何特征进行逐元素加权融合，公式为 f_n = γ_n ⊙ f_box + (1-γ_n) ⊙ f_cls，使模型能按空间位置自适应调整语义与几何信息贡献。\n- Class-Aware Rendering根据类别在训练集中的频率定义优先级 p(c_k)=1/log(1+f_{c_k})，在栅格化时让低频稀有类别覆盖高频类别，避免自行车、拖车等被车辆覆盖。\n- 蒸馏损失采用前景掩码约束的特征级L2距离，仅对ground-truth目标中心高斯热图对应的前景区域计算，并利用通道划分策略保留图像特征自身语义容量。\n- 实验基于nuScenes数据集，使用BEVDepth代码库、ResNet-50骨干、256×704输入和5帧时序输入，在4块NVIDIA A40 GPU上训练，推理时无额外模块。\n- 在nuScenes验证集上，LGD-BEV达到40.2 mAP和51.5 NDS，超过StreamPETR和SOLOFusion等复杂模型，并接近使用CBGS的DistillBEV（40.3 mAP, 51.0 NDS）。\n- 消融实验显示，加入GFM和CaR后学生模型从39.7 mAP/50.4 NDS提升到40.2 mAP/51.5 NDS，验证了两个模块对标签特征质量和学生性能的贡献。\n- 特征通道划分实验表明，当标签特征与图像特征通道比例为1:3时性能最佳，说明图像特征对最终检测贡献更大，标签特征主要起引导作用。\n- 距离分组实验显示，GFM和CaR同时使用可改善近处（≤30m）和远处（≥30m）目标的定位与朝向误差，尤其降低远处目标的mAOE。\n- 方法的主要限制是依赖ground-truth标注质量，若标签含噪声或低质量，性能可能下降；论文未验证半监督或合成标签场景的适应性。"
}