SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving
SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving
Section titled “SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving”⚠️ AI 生成 · 建议对照原文 本页为自动整理的学习笔记;关键数据与引用如需引用,请回查 PDF / 官方版本。
学习档位 中文笔记
类型 文献 · 更新 2026-07-19
中文学习笔记(自动生成,需核验)
Section titled “中文学习笔记(自动生成,需核验)”Topic: spatial-perception · Tier: recent · Year: 2023 · Venue: —
Evidence level: partial · 本地全文: 是 · 建议阅读: ~40 分钟
Paper: https://arxiv.org/abs/2303.09551
Code: —
Generator: grok
为什么值得读
Section titled “为什么值得读”提出多相机输入下的密集3D占用预测框架SurroundOcc,通过2D-3D空间注意力与多尺度体积监督实现更全面的场景几何建模,并设计无需昂贵人工标注的密集占用真值生成管道,弥补检测与深度图在任意形状/长尾类别描述上的不足,对自动驾驶视觉感知与后续语义扩展有直接参考价值。
SurroundOcc用多尺度2D特征经2D-3D可变形注意力提升为3D体积,再3D卷积上采样多级监督,并借助多帧LiDAR拼接+Poisson重建生成密集占用标签,实现多相机密集3D占用预测。
从多相机图像预测周围场景的3D占用(每个体素的占用概率或语义占用),以克服3D目标检测难以覆盖任意形状与无限类别、深度图无法恢复遮挡部分的问题,实现更完整的多相机几何一致3D场景理解。
多相机/环视3D感知基础、体素/体积表示、2D骨干与特征金字塔、可变形注意力/交叉注意力、3D卷积与U-Net式上采样、LiDAR点云处理与Poisson表面重建、nuScenes与SemanticKITTI等数据集及占用/场景补全相关指标。
- 提出SurroundOcc:多相机图像到密集3D占用预测的端到端方法,采用3D体积而非BEV表示以更好保留3D信息。
- 设计2D-3D空间(可变形)注意力,将多尺度图像特征提升到3D体积查询,并用3D卷积交互邻域。
- 构建多尺度占用预测(类U-Net),通过3D反卷积融合与逐级衰减损失监督高低层特征。
- 提出仅用现有3D检测与语义分割标签生成密集占用真值的管道:动态物体与静态场景分别多帧拼接、Poisson重建填洞、体素化+最近邻赋语义,避免昂贵占用标注。
输入多相机图像I → 2D骨干提取多尺度特征X → 对每尺度用2D-3D空间可变形注意力将3D体积查询投影到命中视图并聚合特征得到体积F → 3D反卷积上采样并与更高分辨率体积融合(Yj = Fj + Deconv(Yj-1))→ 各尺度输出占用V并施加衰减加权损失(交叉熵+场景类亲和损失)→ 推理得到最终3D占用。真值生成:按框分离动/静态点→多帧世界坐标系拼接→Poisson重建网格→体素化→NN从稀疏语义体素赋标签。
关键模块和设计取舍
Section titled “关键模块和设计取舍”3D体积查询(vs BEV,更好保留高度与3D结构);投影+可变形注意力仅聚合命中视图特征(避免等权平均被遮挡/模糊视图影响);多尺度2D-3D注意力+3D deconv融合(低层细节+高层语义);衰减损失权重αj=1/2^j(重视高分辨率);动静态分开拼接+Poisson填洞+NN语义(解决单帧稀疏、运动物体与孔洞问题)。取舍:用体积而非BEV/深度以支持遮挡补全与多相机一致性;用现有标签生成密集GT而非昂贵人工标注;监督用交叉熵+MonoScene类亲和损失。
数据集、实验设置与指标
Section titled “数据集、实验设置与指标”主要在nuScenes上训练/评估(多相机环视),并在SemanticKITTI上验证(虽非专为单目设计,但单目3D语义场景补全达SOTA)。指标包括场景补全SC IoU与语义场景补全SSC mIoU,以及各类别IoU(表格列出barrier、bicycle、bus、car等)。实验设置与完整超参细节在摘录中未完全展开。
主要结果(需原文证据)
Section titled “主要结果(需原文证据)”摘录表格显示对比方法:MonoScene SC IoU 23.96 / SSC mIoU 7.31;Atlas 28.66 / 15.00;BEVFormer 30.50 / 16.75;TPVFormer 11.51 / 11.66;TPVFormer* 30.86 / 17.10(后续方法与SurroundOcc具体数字因表格截断待来源核验)。文中称在nuScenes上定量与可视化均展示优越性,且在SemanticKITTI单目3D语义场景补全基准达到state-of-the-art;密集真值使预测更密。
局限、失败场景与适用边界
Section titled “局限、失败场景与适用边界”摘录未系统讨论失败场景;潜在边界包括:依赖LiDAR点与检测/分割标签生成GT(表面占用为主,内部可能不完整);Poisson与NN可能引入伪影;多帧拼接依赖准确位姿与物体跟踪;方法面向多相机,单目为额外验证;完整消融与失败案例待来源核验。
与前序 / 同期 / 后续方法的关系
Section titled “与前序 / 同期 / 后续方法的关系”相对MonoScene(单目占用,简单后处理融合多相机效果差)与TPVFormer(稀疏LiDAR监督导致稀疏预测)更强调密集监督与多相机体积注意力;相对BEVFormer等BEV检测/表示,改用体积占用以更细粒度与遮挡建模;相对深度估计(仅最近表面、难恢复遮挡)与室内多视图重建(Atlas、NeuralRecon等),适配户外多相机自动驾驶;可扩展到语义分割等下游。
官方代码与复现建议
Section titled “官方代码与复现建议”推荐阅读顺序
Section titled “推荐阅读顺序”先读Abstract与Introduction(问题与动机)→ Section 3 Approach(问题定义、2D-3D注意力、多尺度预测)→ Section 4 Dense Occupancy Ground Truth(拼接-Poisson-NN管道)→ Related Work对比→ 实验表格与可视化(摘录截断处需补全文)→ 结论与代码。
- Q: 为什么作者主张3D占用优于深度图或纯3D检测? A: 占用可描述任意形状与更多类别、天然多相机几何一致、能利用语义线索预测遮挡区域,并易于扩展到语义分割等下游;深度仅最近表面,检测难覆盖长尾。
- Q: 2D-3D空间注意力如何工作?与简单反投影平均有何不同? A: 构建3D体积查询,投影到命中2D视图后用可变形注意力采样并加权聚合特征;避免不同视图等权平均(遮挡/模糊视图贡献不应相同)。
- Q: 密集占用真值如何生成?关键步骤是什么? A: 动静态物体/场景分别用检测框切分并多帧世界坐标系拼接→计算法向后Poisson重建填洞得网格→体素化→用NN从稀疏语义体素赋标签。
- Q: 多尺度监督如何加权?为什么? A: 用衰减权重αj=1/2^j,高分辨率预测更重要,同时保留高低层特征能力。
- Q: SurroundOcc与TPVFormer的主要区别? A: TPVFormer用稀疏LiDAR监督导致稀疏预测;SurroundOcc设计密集真值生成管道并采用多尺度体积监督,预测更密。
- page 1 Abstract: we propose a SurroundOcc method to predict the 3D occupancy with multi-camera images. We first extract multi-scale features for each image and adopt spatial 2D-3D attention to lift them to the 3D volume space. Then we apply 3D convolutions to progressively upsample the volume features and impose supervision on multiple levels. To obtain dense occupancy prediction, we design a pipeline to generate dense occupancy ground truth without expansive occupancy annotations. Specifically, we fuse multi-frame LiDAR scans of dynamic objects and static scenes separately. Then we adopt Poisson Reconstruction to fill the holes and voxelize the mesh to get dense occupancy labels.
- page 1 Abstract / Code: Code and dataset are available at https://github.com/weiyithu/SurroundOcc.
- page 3 Section 3.1 Problem Formulation: V = G(I 1 , I 2 , · · · I N ) where G is an neural network and V ∈ RH×W ×Z is the 3D occupancy. The value of V is between 0 and 1, representing the occupied probability of the grids.
- page 4 Section 3.3 / Eq.2: instead of 2D BEV queries, we build 3D volume queries to further reserve 3D space information. … DeformAttn … Fp = 1/|Vhit| ∑ i∈Vhit DeformAttn(Qp , P(qp , i), Xi )
- page 4 Section 3.4: Yj = Fj + Deconv(Yj−1 ) … we use a decayed loss weight αj = 1/2^j for jth level supervision.
- page 5 Section 4 / Figure 4 caption: We first traverse all frames to stitch the multi-frame LiDAR points of dynamic objects and static scenes separately, and then merge them into a complete scene. Subsequently, we employ Poisson Reconstruction to densify the points and voxelize the resulting mesh to obtain a dense 3D occupancy. Finally, we use the Nearest Neighbor (NN) algorithm to assign semantic labels to dense voxels.
- page 6 Table (partial): MonoScene [8] SC IoU 23.96 SSC mIoU 7.31 … Atlas [37] 28.66 15.00 … BEVFormer [29] 30.50 16.75 … TPVFormer [22] 11.51 11.66 … TPVFormer* 30.86 17.10
- page 2 Introduction / Related: TPVFormer [22] uses sparse LiDAR points as supervision, which results in sparse occupancy prediction. … Although our method is not designed for the monocular setting, it achieves state-of-the-art performance on the monocular 3D semantic scene completion benchmark.
Discovery evidence
Section titled “Discovery evidence”- topic:
spatial-perception - sources:
asta,openalex - retrieved_at: 2026-07-20
- query: Find foundational and recent research papers for the topic «空间感知» (spatial-perception). Prefer peer-reviewed or widely cited work with clear method contributions. Include open-source code when available. Exclude pure survey spam unless highly cited. Core concepts: depth estimation, occupancy, 3D reconstruction. Search facets: occupancy prediction autonomous driving; depth estimation multi-camera BEV; 3D scene reconstruction driving. Relevant venues include: CVPR, ICCV, ICRA. Return papers with r
- corpus_id:
257557568 - doi:
10.1109/iccv51070.2023.01986 - relevance_score:
0.916510157707931 - score_total: 67
- suggested_tier:
recent
Relevance
Section titled “Relevance”(no prose relevance explanation — numeric score only or HTTP source)
Snippets
Section titled “Snippets”(no snippet evidence in candidate pool)
延伸解读与背景补充
Section titled “延伸解读与背景补充”生成:2026-07-21 · 来源条数 2 · 模型
heuristic· 需人工核验数字
围绕「SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving」的核心问题与动机(待结合全文核验)。
方法要点补强
Section titled “方法要点补强”- 见原文方法章节;以下为基于摘要/摘录的要点提示。
- SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving Yi Wei1,2 ,* Linqing Zhao3 ,* W…
与相近工作的关系
Section titled “与相近工作的关系”与相近工作的关系待核验;请对照 related work。
- 勿仅凭摘要推断未给出的数值指标。
- 这篇工作的输入/输出表示是什么?(SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving)
- 训练目标与评测协议各是什么?
- 主要失败模式或局限是什么?
外部解读索引
Section titled “外部解读索引”- [2303.09551] SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving — 2026-07-21 — 官方摘要/二次页面(自动抓取)
- [2303.09551] SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving — 2026-07-21 — 官方摘要/二次页面(自动抓取)
方法结构(重绘)
Section titled “方法结构(重绘)”flowchart LR A["输入 / 观测"] --> B["表示 / 编码"] B --> C["推理 / 解码"] C --> D["输出 / 动作或检测"] %% method sketch for: SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving方法结构示意(重绘;细节以原论文为准,待 PDF 核验)。
论文摘录图/表
Section titled “论文摘录图/表”
来源:原论文约 p.1(arch);学习用途摘录。

来源:原论文约 p.7(table);学习用途摘录。
英文自动分析(可折叠)
Section titled “英文自动分析(可折叠)”展开英文 Paper Card / AI deep analysis
Paper Card
Section titled “Paper Card”| Field | Content |
|---|---|
| Year | 2023 |
| Authors | Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, Jiwen Lu |
| arXiv | 2303.09551 |
| DOI | 10.1109/iccv51070.2023.01986 |
| Topics | spatial-perception, scene-representation-memory |
| Paper | https://arxiv.org/abs/2303.09551 |
原文摘录与素材(可折叠)
Section titled “原文摘录与素材(可折叠)”展开 Extract / Selections / Local assets
Selections
Section titled “Selections”spatial-perception: tier=recent score=50 — auto refresh 2026-07-19 sources=openalex | promoted watch->recent for coverage fillscene-representation-memory: tier=recent rank=1 score=85 — SurroundOcc multi-camera 3D occupancy — dense scene memory
Extract excerpt
Section titled “Extract excerpt”(no PDF text available; metadata-only card)