{
  "paper_id": "b94dd3d28a09809febd833d1924d949c",
  "paper_title": "OneBEV++: Towards Unifying Bird's-Eye-View Semantic Mapping with Panoramas",
  "paper_authors": [
    "Jiale Wei",
    "Zhifeng Teng",
    "Fei Teng",
    "Junwei Zheng",
    "Ruiping Liu",
    "Yufan Chen",
    "Jie Hu",
    "Kailun Yang",
    "Jiaming Zhang",
    "Rainer Stiefelhagen"
  ],
  "publication_date": "2026",
  "publication_venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
  "abstract": "Bird’s-Eye-View (BEV) perception, crucial for holistic scene understanding, faces limitations with narrow field-of-view sensors (indoors) and complex multi-camera setups (outdoors). To address these limitations and enable omni-range perception with a single image, we propose a new pipeline called panoramic-to-BEV semantic mapping. We extend existing household-related and driving-scene datasets to create four benchmarks for Panoramic BEV in both indoor and outdoor settings. As ground-breaking solutions, the 360BEV (indoor) and OneBEV (outdoor) methods have shown promise in their respective domains. Moreover, we propose a unified framework OneBEV++, generating BEV semantic maps from a single panoramic image across all environments. Extensive experiments on our four panoramic datasets demonstrate that OneBEV++ outperforms previous methods with consistent accuracy gains and improved efficiency. By simplifying complexity while enhancing performance, OneBEV++ enables scalable applications across diverse environments from outdoor autonomous vehicles to indoor robots, paving the way for more holistic cross-environment autonomous systems. Data and code will be available at <https://github.com/JialeWei/OneBEV>.\n",
  "keywords": [
    "Omni-range Image Processing",
    "View Transformation",
    "Semantic Mapping",
    "Scene Understanding"
  ],
  "paper_institutions": [
    "Karlsruhe Institute of Technology (KIT)",
    "Hunan University",
    "XYZ Robotics GmbH"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "deepseek-v4-flash",
  "prompt_template": "simple",
  "extracted_at": "2026-08-18T10:32:07.702959Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n本文针对鸟瞰图（BEV）语义映射中窄视场传感器和复杂多相机设置带来的局限，提出以单张全景图像实现跨室内外环境的全景到BEV语义映射。现有工作如360BEV（室内）和OneBEV（室外）分别在各自领域有效，但缺乏跨域统一能力；OneBEV++在二者基础上提出统一框架，通过柱状径向投影（CRP）建立BEV坐标与全景图像列之间的几何对应（核心公式为方位角 φ_i = atan2(-x_i, y_i)，经2π调整后得到列索引 c_i = φ_i/(2π)），并设计特征选择块（FSB）结合最大池化与状态空间建模，集成到优化的MVTv2视图变换模块中。数据集方面，扩展并标准化了Stanford-360、Matterport-360（室内）和NuScenes-360、DeepAccident-360（室外）四个基准，覆盖真实与合成、室内与室外场景，以验证跨环境泛化能力；实现上采用VMamba-T骨干，MVTv2配置2层、2个采样位置、每位置2个偏移点，嵌入维度128，4块A100训练50轮，室内输入512×1024、室外180×3200，BEV输出200×200，感知范围室内±5m、室外±50m。以mIoU为指标，OneBEV++在NuScenes-360上达56.3%、DeepAccident-360上38.2%、Stanford-360上22.0%、Matterport-360上28.8%（val）/24.3%（test），同时相比分别使用专用模型降低46% FLOPs和7%参数量，在效率和精度上均优于此前方法。进一步探索包括MVT/CRP组件消融、视场角影响和拼接鲁棒性分析；其适用场景涵盖户外自动驾驶与室内机器人，约束包括无深度条件下的投影精度上限、拼接伪影以及缺失视角时的性能下降。\n\n### 关键要点\n\n- OneBEV++提出统一的全景到BEV语义映射框架，仅需单张全景图像即可同时处理室内和室外环境，无需针对特定领域重新设计。\n- 核心组件Columnar Radial Projection (CRP)通过方位角计算将BEV查询点映射到全景图像列，公式为φ_i=atan2(-x_i,y_i)，并调整到[0,2π)后得到列索引c_i=φ_i/(2π)。\n- MVTv2模块包含查询采样、特征选择与状态空间建模、迭代增强三个关键操作，其中Feature Selection Block (FSB)使用最大池化提取显著特征，再交由Vanilla VSS Block进行空间建模。\n- 数据集方面，作者扩展了Stanford-360和Matterport-360两个室内基准，以及NuScenes-360和DeepAccident-360两个室外基准，并统一了标注协议和评估指标。\n- 实现设置上，OneBEV++采用VMamba-T骨干，MVTv2配置2层、2个采样位置和2个偏移点，使用4块A100 GPU训练50轮，初始学习率5×10^-5，AdamW优化器。\n- 在室外数据集上，OneBEV++在NuScenes-360达到56.3% mIoU，在DeepAccident-360达到38.2% mIoU，分别比OneBEV提高5.2%和2.1%。\n- 在室内数据集上，OneBEV++在Stanford-360达到22.0% mIoU，在Matterport-360验证集和测试集分别达到28.8%和24.3% mIoU。\n- 效率方面，OneBEV++仅28.9M参数和88G FLOPs（室内）/94.1G FLOPs（室外），相比分别使用360BEV和OneBEV的专用模型降低46% FLOPs和7%参数量，同时保持28.15 FPS。\n- CRP消融实验表明，学习到的采样偏移贡献最大（移除后mIoU下降8.29），几何列先验比完全学习的采样高1.69 mIoU，而使用真值深度的IRP上界为45.80 mIoU，说明无深度投影存在固有精度上限。\n- 视场角与鲁棒性分析显示，360°输入相比60°带来+23.89 mIoU提升；在拼接扰动中，外观级扰动（接缝偏移、曝光失配）影响很小，而几何级扰动（偏航偏差、缺失视角）会导致明显下降，但OneBEV++仍保持对OneBEV约+5 mIoU的优势。"
}