{
  "paper_id": "da0dafddbd36994f740309f11ca4eb61",
  "paper_title": "TiG-BEV: Multi-view BEV 3D Object Detection via Target Inner-Geometry Learning",
  "paper_authors": [
    "Peixiang Huang",
    "Li Liu",
    "Renrui Zhang",
    "Song Zhang",
    "Xinli Xu",
    "Baichao Wang",
    "Guoyi Liu"
  ],
  "publication_date": "",
  "publication_venue": "unknown",
  "abstract": "To achieve accurate multi-view 3D object detection, existing methods propose to benefit camera-based detectors with spatial cues provided by the LiDAR modality, e.g., dense depth supervision and bird-eye-view (BEV) feature distillation. However, they directly conduct point-to-point mimicking from LiDAR to camera, which neglects the inner-geometry of foreground targets and suffers from the modal gap between 2D-3D features. In this paper, we propose the learning scheme of Target Inner-Geometry from the LiDAR modality into camera-based BEV detectors for both dense depth and BEV features, termed as TiG-BEV. First, we introduce an inner-depth supervision module to learn the low-level relative depth relations between different foreground pixels. This enables the camera-based detector to better understand the object-wise spatial structures. Second, we design an inner-feature BEV distillation module to imitate the high-level semantics of different keypoints within foreground targets. To further alleviate the BEV feature gap between two modalities, we adopt both interchannel and inter-keypoint distillation for feature-similarity modeling. With our target inner-geometry learning, TiG-BEV can effectively boost student models by different margins on nuScenes val set, e.g., +2.3% NDS and +2.4% mAP for BEVDepth, along with +9.1% NDS and +10.3% mAP for BEVDet without CBGS. Code will be available at <https://github.com/ADLab3Ds/TiG-BEV>.\n",
  "keywords": [
    "3D object detection",
    "BEV",
    "knowledge distillation",
    "depth estimation",
    "LiDAR-to-camera learning",
    "nuScenes"
  ],
  "paper_institutions": [
    "Peking University",
    "NIO",
    "The Chinese University of Hong Kong",
    "University of Chinese Academy of Sciences"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "deepseek-v4-flash",
  "prompt_template": "simple",
  "extracted_at": "2026-08-18T10:30:13.287243Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n本文针对多视角BEV 3D目标检测中相机与LiDAR模态差异以及现有方法忽略前景目标内部几何结构的问题，提出TiG-BEV目标内部几何学习方案。相关工作包括基于稠密深度监督（如BEVDepth）和BEV特征蒸馏（如BEVDistill）的LiDAR到相机迁移，但这些方法逐点模仿或逐通道对齐，受模态间隙影响。TiG-BEV包含两个核心模块：内深度监督模块利用连续深度值并自适应选取参考点，计算前景目标内像素的相对深度，核心公式为 d_j(x,y)=Σ_k d[k]·S_j[k](x,y)，从而学习低层目标空间结构；内特征BEV蒸馏模块在BEV空间中采样前景关键点，分别进行通道间和关键点间相似性蒸馏，缓解跨模态语义差距。实验在nuScenes数据集上进行，采用CenterPoint作为LiDAR教师，BEVDepth、BEVDet、BEVDet4D作为相机学生，在8块NVIDIA A100 GPU上按官方配置训练；在nuScenes val上，BEVDepth提升+2.3% NDS和+2.4% mAP，BEVDet（无CBGS）提升+9.1% NDS和+10.3% mAP，消融实验验证两个模块互补。未来工作将探索同时提升相机与LiDAR模态的统一多模态感知策略；实际场景中该方法适用于自动驾驶等低成本视觉感知，训练时需要LiDAR教师和真值框，推理时仅需多视角图像，且受LiDAR点云稀疏性和模态差异约束。\n\n### 关键要点\n\n- TiG-BEV提出从LiDAR模态向相机BEV检测器迁移“目标内部几何”知识，包含内深度监督和内特征BEV蒸馏两个互补模块。\n- 内深度监督在现有绝对深度监督基础上，利用LiDAR投影深度构造前景目标内像素的连续相对深度，帮助模型理解物体内部空间结构。\n- 内深度监督采用自适应深度参考点，选择预测误差最小的像素作为参考，公式为 (x_r,y_r)=argmin(S_j^gt(x,y)-S_j(x,y))，以适应不同形状和遮挡目标。\n- 内特征BEV蒸馏在BEV空间对每个前景目标均匀采样N个关键点，并通过双线性插值提取相机和LiDAR的关键点特征。\n- 蒸馏损失包含通道间相似性蒸馏 L_bev^IC 和关键点间相似性蒸馏 L_bev^IK，分别建模通道多样性和关键点语义相关性，避免逐点严格对齐。\n- 总体损失为 L_TiG = L_det + L_depth^A + L_depth^R + L_bev^IC + L_bev^IK，其中L_depth^R为内深度L2损失，L_bev为两种蒸馏损失之和。\n- 实验采用nuScenes数据集，CenterPoint作为预训练LiDAR教师，BEVDepth、BEVDet、BEVDet4D作为学生，在8块NVIDIA A100 GPU上基于BEVDet代码库实现。\n- 在nuScenes val集上，TiG-BEV使BEVDepth提升+2.3% NDS和+2.4% mAP，使BEVDet（无CBGS）提升+9.1% NDS和+10.3% mAP。\n- 与BEVDistill相比，在相同设置下TiG-BEV在更强基线上仍获得更高提升，表明目标内部几何学习优于前景引导的稠密蒸馏。\n- 消融实验显示，仅内深度监督带来+1.0% mAP和+0.9% NDS，仅内特征蒸馏带来+3.0% mAP和+2.3% NDS，两者结合达到+3.7% mAP和+2.7% NDS。\n- 在不同2D骨干（ResNet-18/50/101）和单帧/多帧设置下，TiG-BEV均带来一致提升，mAP提升+3.4%至+5.8%，NDS提升+2.5%至+5.0%。\n- 推理时相机检测器仅需多视角图像，不依赖LiDAR或教师模型；训练时需要LiDAR点云、真值3D框和预训练LiDAR教师，适用于自动驾驶等低成本视觉感知场景。"
}