{
  "paper_id": "1b84254f626d8505e55c04f2a08cc773",
  "paper_title": "SSC-RS: Elevate LiDAR Semantic Scene Completion with Representation Separation and BEV Fusion",
  "paper_authors": [
    "Jianbiao Mei",
    "Yu Yang",
    "Mengmeng Wang",
    "Tianxin Huang",
    "Xuemeng Yang",
    "Yong Liu"
  ],
  "publication_date": "",
  "publication_venue": "2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)",
  "abstract": "Semantic scene completion (SSC) jointly predicts the semantics and geometry of the entire 3D scene, which plays an essential role in 3D scene understanding for autonomous driving systems. SSC has achieved rapid progress with the help of semantic context in segmentation. However, how to effectively exploit the relationships between the semantic context in semantic segmentation and geometric structure in scene completion remains under exploration. In this paper, we propose to solve outdoor SSC from the perspective of representation separation and BEV fusion. Specifically, we present the network, named SSC-RS, which uses separate branches with deep supervision to explicitly disentangle the learning procedure of the semantic and geometric representations. And a BEV fusion network equipped with the proposed Adaptive Representation Fusion (ARF) module is presented to aggregate the multi-scale features effectively and efficiently. Due to the low computational burden and powerful representation ability, our model has good generality while running in real-time. Extensive experiments on SemanticKITTI demonstrate our SSC-RS achieves state-of-the-art performance. Code is available at <https://github.com/Jieqianyu/SSC-RS.git>.\n",
  "keywords": [
    "语义场景补全",
    "LiDAR",
    "BEV融合",
    "表示分离",
    "自动驾驶"
  ],
  "paper_institutions": [],
  "output_language": "zh",
  "provider": "doubao",
  "model": "deepseek-v4-flash",
  "prompt_template": "simple",
  "extracted_at": "2026-08-18T10:30:15.539686Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n本文针对自动驾驶中 LiDAR 语义场景补全（SSC）任务，提出从表示分离与BEV融合角度解决语义上下文与几何结构关系利用不足的问题。现有方法多以混合或半混合方式同时处理语义与几何，SSC-RS 则显式解耦两类表示：语义分支采用稀疏3D CNN编码多尺度语义上下文，补全分支采用轻量稠密3D CNN获取多尺度几何结构，并对两个分支施加深度监督；随后通过BEV融合网络中的自适应表示融合（ARF）模块聚合多尺度特征。ARF模块对来自前级、语义分支和补全分支的特征分别计算通道注意力并加权求和，再经1×1卷积得到融合特征，其核心机制可表示为 F_f = φ{σ[MLP(AvgPool(F_prev))]×F_prev + σ[MLP(AvgPool(F_sem))]×F_sem + σ[MLP(AvgPool(F_com))]×F_com}。在SemanticKITTI数据集上，输入范围设为[0~51.2m, -25.6~25.6m, -2~4.4m]，体素分辨率0.2m，空间分辨率256×256×32，使用Adam优化器训练40轮；实验表明该方法在官方测试集上达到IoU 59.7%、mIoU 24.2%，并以16.7 fps实时运行，在补全指标上超越已发表方法，mIoU排名第二。消融实验验证了语义表示、几何表示、ARF模块和多尺度监督各自的有效性。未来工作将关注小物体的局部几何学习，并扩展到室内和单目等更多场景。该方法适用于自动驾驶中因遮挡或远距离导致的点云稀疏与不完整场景，其约束包括对小型物体性能仍有不足，且未专门设计局部几何损失。\n\n### 关键要点\n\n- SSC-RS 将语义上下文与几何结构的学习过程显式解耦，分别通过语义分支和补全分支进行表示学习，并在BEV中融合。\n- 语义分支由体素化层和三个稀疏编码器块组成，使用稀疏3D CNN编码多尺度语义上下文，参数量仅1.45M。\n- 补全分支以占用体素为输入，采用稠密3D卷积和最大池化获取多尺度几何结构，参数量仅0.31M，计算开销为7.93G MACs。\n- 两个分支在训练时均使用多尺度深度监督，语义分支采用Lovasz损失和交叉熵损失，补全分支采用Lovasz损失和二元交叉熵损失，推理时辅助头被移除。\n- BEV融合网络采用2D U-Net结构，将稀疏语义特征通过聚合函数投影为BEV特征，将稠密几何特征沿z轴堆叠并降维，从而高效融合多尺度表示。\n- ARF模块对前级、语义和补全特征分别进行通道注意力加权，再求和并经过1×1卷积，能够选择不同来源的有效线索。\n- 总训练损失为 L_total = 3·L_bev + L_s + L_c，其中L_bev对最终BEV预测使用Lovasz损失和交叉熵损失。\n- 在SemanticKITTI上，序列00-10除08外用于训练，08用于验证，其余用于测试；输入范围为[0~51.2m, -25.6~25.6m, -2~4.4m]，体素分辨率0.2m，空间分辨率256×256×32。\n- 官方隐藏测试集上，SSC-RS达到IoU 59.7%、mIoU 24.2%，运行速度16.7 fps，在IoU上超过所有已发表方法，mIoU排名第二。\n- 消融实验显示，移除语义分支使mIoU从24.8%降至21.9%，移除多尺度监督使mIoU下降2.1%，ARF模块带来0.4% mIoU和0.2% IoU提升。\n- 与SSA-SC相比，SSC-RS在IoU上提高0.9%、mIoU上提高0.7%；与S3CNet相比，IoU提高14.1%且速度快约14倍，但在小物体上仍不如S3CNet。\n- 当前方法对行人、自行车等小物体的完成效果有限，未来计划引入局部几何学习，并扩展到室内和单目场景。"
}