PUMA
PUMA 把最近 N 次掃描累積成局部點雲，以 Poisson 表面重建生成三角網格，並依頂點密度修剪 10% 低支持頂點，移除 Poisson 在無資料處外插的表面；新掃描以射線投射求與網格三角面的交點作為對應，進行點對面 ICP（frame-to-mesh）。局部網格每 M 次掃描併入全域網格，全域網格不參與估計，也未實作迴圈修正。
本頁內容
LiDAR odometry and mapping that builds sliding-window Poisson meshes, trims low-density vertices, and registers scans to the mesh with ray-casting-based point-to-plane ICP.
技術屬性
欄位內容為文獻擷取紀錄的原文用語（英文），以原文為據；「未查證」表示本研究尚未讀到該資訊，不代表該方法不具備此能力。
| 感測輸入 | 3D LiDAR |
|---|---|
| 原文測試平台 | vehicle、simulation |
| 狀態估計 | iterative frame-to-mesh point-to-plane ICP with Huber kernel, initialised with the previous pose increment; during the first N = 30 scans, before a mesh exists, standard point-to-plane ICP is used (Sec. III-A, III-C, IV-D) |
| 資料關聯 | rays from the current sensor origin through every scan point are intersected with the local triangle mesh (Embree); the hit point and triangle normal form the correspondence, and pairs farther apart than 1 m are rejected (Sec. III-A, IV) |
| 時間表示 | discrete poses |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | none (listed as future work, Sec. V) |
| 全域最佳化 | none |
| 地圖表示 | local Poisson surface reconstruction mesh (octree depth 10) rebuilt from the last N = 30 scans after every registered scan, with the 10% lowest-density vertices trimmed; global mesh aggregated every M = 30 scans with duplicate-triangle removal and used only for visualisation and output (Sec. III-B, III-C, IV) |
| 先驗資訊 | none |
| 可輸出幾何 | global triangle mesh |
| 計算需求 | CPU only (Intel Xeon W-2145, 8 cores at 3.70 GHz, 32 GB RAM; built on Open3D, Intel Embree for ray-triangle queries); per scan about 45 ms preprocessing and normals, 500 ms scan matching and about 5 s Poisson meshing, so online operation is infeasible (Sec. IV, IV-F) |
使用設備
原文使用的感測器、運算硬體與載具（equipment）。型號保留原文寫法，連結到設備頁中同一型號的歸併名稱；角色依原文用途分為方法輸入、資料集感測器、執行運算平台、參考或真值量測（reference or ground truth）與比較對象設備。
| 類別 | 型號（原文寫法） | 角色 | 資料集 | 原文規格 | 出處 |
|---|---|---|---|---|---|
| LiDAR | rotating 3D LiDAR of the KITTI odometry benchmark (model not named in the paper) | 資料集感測器 | KITTI Odometry | real-world odometry, registration and memory experiments on KITTI sequences 00-10 | (Vizzo et al., 2021, Sec. IV-A; Tables II-III) |
| LiDAR | virtual 64-beam LiDAR sensor model (Mai City scans) | 資料集感測器 | Mai City | synthetic scans of a CAD urban scene; a 64-beam sensor is implied by the statement that the ground-truth sampling used 320 beams instead of 64 (Sec. IV-A) | (Vizzo et al., 2021, Sec. IV-A) |
| LiDAR | virtual 320-beam LiDAR sensor model (Mai City ground truth) | 參考或真值量測 | Mai City | samples the CAD model with 320 beams to give a 62.5-million-point ground-truth cloud of the observable surfaces | (Vizzo et al., 2021, Sec. IV-A; Fig. 3) |
| 運算硬體 | Intel Xeon W-2145 | 執行運算平台 | 未標示 | 8 cores at 3.70 GHz, 32 GB RAM; algorithm runs entirely on CPU | (Vizzo et al., 2021, Sec. IV) |
作者報告的優勢與限制
優勢
- On synthetic Mai City data with ground-truth poses (d = 0.03 m), the mesh map scored Chamfer 0.05, precision 93.28, recall 98.69 and F-score 95.91, versus TSDF and surfel maps built with the same poses (Table I).
- Mesh maps need far less memory than point clouds or surfels and about the same as TSDF-extracted meshes on KITTI 04 and 07; final mesh sizes 72 MB and 306 MB (Sec. IV-C, Fig. 4).
- Ray-casting association was more accurate and faster than mesh vertex sampling at all three mesh resolutions (Table III).
- On KITTI 00-10 the ray-casting variant reached 1.55% average relative translational error and 0.74 deg/100 m, versus 2.93% and 0.92 deg/100 m for frame-to-frame SuMa under the same normal computation (Table II).
限制
- Meshing takes about 5 s per scan on CPU, making online operation infeasible (Sec. IV-F).
- No loop closure; proposed as future work (Sec. V).
- Watertight Poisson assumption requires density trimming for open outdoor scenes (Sec. III-B).
- Ray-casting association does not handle large rotational motions properly and needs a good initial estimate (Sec. III-A).
- Only horizontally mounted LiDARs are considered; upward-looking profiler scanners are not examined (Sec. III-B, footnote 1).
- Range-image cross-product normals are sometimes less accurate than PCA normals (Sec. III-A), and the authors note that imposing them on all methods generally lowers the performance of normal-based metrics (Sec. IV-D).
- SLAMesh authors report PUMA meshes are multi-layered and slightly warped at wall tops (Ruan et al. 2023, Sec. IV).
營建工程相關證據
僅在合成城市資料（Mai City）與 KITTI 驗證，未涉及營建；密度修剪會移除低觀測區域，對構件覆蓋完整度的影響未研究（推論）。
原文驗證環境：模擬、公開基準
報告的性能數據
以下是原文作者報告的性能數值（author-reported results），不是本研究重新量測的結果。每張圖只並列同一個比較組（comparison group，同一張表、同一組實驗設定）內的方法；不同比較組之間的數值不可直接比較，也不構成排名。
本方法共出現在 22 個比較組，合計 128 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 18 組列在最後，並連到性能比較頁。
Vizzo et al., 2021 · Table II 本方法 26 筆
表格設定（擷取紀錄原文）：KITTI odometry training sequences 00-10; relative errors averaged over 100-800 m segments; all methods share the range-image normals and Huber loss; Map None = frame-to-frame, Map Point cloud = frame-to-model on the last N scans; DA = data association (NN nearest neighbour, Proj. projective, RC ray casting). Per-sequence rotational errors omitted to respect the row cap; only the rotational average is kept (Vizzo et al., 2021, Table II)
relative translational error (%)，KITTI Odometry · average of 00-10
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Vizzo et al., 2021 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Vizzo et al., 2021, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| point-to-point ICP [3] (map: None, DA: NN) | 4.86% | (Vizzo et al., 2021, Table II) |
| point-to-plane ICP [32] (map: None, DA: NN) | 7.6% | (Vizzo et al., 2021, Table II) |
| GICP [33] (map: None, DA: NN) | 14.35% | (Vizzo et al., 2021, Table II) |
| SuMa [1] (map: None, DA: Proj.) | 2.93% | (Vizzo et al., 2021, Table II) |
| point-to-point ICP [3] (map: Point cloud, DA: NN) | 29.98% | (Vizzo et al., 2021, Table II) |
| point-to-plane ICP [32] (map: Point cloud, DA: NN) | 18.92% | (Vizzo et al., 2021, Table II) |
| GICP [33] (map: Point cloud, DA: NN) | 20.43% | (Vizzo et al., 2021, Table II) |
| Ours (Δtree = 10) (map: Mesh, DA: NN)本方法 | 2.15% | (Vizzo et al., 2021, Table II) |
| Ours (Δtree = 10) (map: Mesh, DA: RC)本方法原文提出 | 1.55% | (Vizzo et al., 2021, Table II) |
Ruan et al., 2023 · Table II 本方法 13 筆
表格設定（擷取紀錄原文）：KITTI odometry 00-10, relative translation error; baseline values imported from their published papers (not rerun); SLAMesh without loop closure; w/o Comb. and w/o P2Mesh are ablations (Ruan et al., 2023, Table II)
relative translation error (%)，KITTI odometry · 00
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Ruan et al., 2023 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Ruan et al., 2023, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOAM | 0.78% | (Ruan et al., 2023, Table II) |
| A-LOAM | 0.97% | (Ruan et al., 2023, Table II) |
| Suma | 0.7% | (Ruan et al., 2023, Table II) |
| Suma++ | 0.64% | (Ruan et al., 2023, Table II) |
| Litamin2 | 0.7% | (Ruan et al., 2023, Table II) |
| Puma本方法 | 1.46% | (Ruan et al., 2023, Table II) |
| SLAMesh (Ours) Full原文提出 | 0.77% | (Ruan et al., 2023, Table II) |
| SLAMesh w/o Comb.原文提出 | 0.96% | (Ruan et al., 2023, Table II) |
| SLAMesh w/o P2Mesh原文提出 | 0.7% | (Ruan et al., 2023, Table II) |
Zhu et al., 2025 · Table V 本方法 10 筆
表格設定（擷取紀錄原文）：Mesh quality with ground-truth poses for all methods, voxel size 0.1 m, settings of SHINE-Mapping; distances in cm; completion ratio and F-score in % at 10 cm (Mai City) and 20 cm (Newer College) (Zhu et al., 2025, Table V)
Comp. (cm), completion，Mai City · Mai City
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Zhu et al., 2025 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Zhu et al., 2025, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| VDB Fusion [28] | 6.9 cm | (Zhu et al., 2025, Table V (VoR)) |
| Puma [13]本方法 | 32 cm | (Zhu et al., 2025, Table V (VoR)) |
| SHINE-Mapping [30] | 3.2 cm | (Zhu et al., 2025, Table V (VoR)) |
| SLAMesh [14] | 7.5 cm | (Zhu et al., 2025, Table V (VoR)) |
| Ours原文提出 | 2.5 cm | (Zhu et al., 2025, Table V (VoR)) |
Vizzo et al., 2021 · Table III 本方法 9 筆
表格設定（擷取紀錄原文）：Registration of every scan to the local mesh on the full KITTI training sequences; Mesh vertex-sampling = standard point-to-plane ICP on sampled mesh vertices (nearest neighbours) versus Mesh ray-casting (proposed); Poisson octree depth 8, 9, 10 give about 30K, 100K, 300K vertices; relative errors over 100-800 m (Vizzo et al., 2021, Table III)
terr, relative translational error (%)，KITTI Odometry · KITTI training sequences 00-10; Δtree = 8 (30K vertices)
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Vizzo et al., 2021 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Vizzo et al., 2021, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Mesh vertex-sampling (point-to-plane ICP on sampled mesh vertices, nearest neighbours) | 2.82% | (Vizzo et al., 2021, Table III) |
| Mesh ray-casting (proposed)本方法原文提出 | 1.73% | (Vizzo et al., 2021, Table III) |
其他比較組
列出其餘 18 個比較組
- Deng et al., 2023 · Table 1
- Deng et al., 2023 · Table 5
- Pan et al., 2024 · Table XI
- Zhu et al., 2025 · Table III
- Zhu et al., 2025 · Table VIII
- Zhong et al., 2023 · Table II
- Zhong et al., 2023 · Table III
- Deng et al., 2023 · Table 3
- Vizzo et al., 2021 · Table I
- Ruan et al., 2023 · Table I
- Vizzo et al., 2021 · Text Sec.IV-F
- Vizzo et al., 2021 · Text Sec.IV-C (Fig. 4 labels)
- Zhu et al., 2025 · Table I
- Zhu et al., 2025 · Table IV
- Ruan et al., 2023 · Text Sec. IV-D
- Zhang et al., 2024a · Table 8
- Zhang et al., 2024a · Table 9
- Zhu et al., 2025 · Table II
來源
Vizzo et al., 2021
（2021）Poisson Surface Reconstruction for LiDAR Odometry and Mapping2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 5624-5630
DOI 10.1109/icra48506.2021.9562069程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 程式碼釋出：PRBonn/puma https://github.com/PRBonn/puma
程式碼：https://github.com/PRBonn/puma（授權：MIT (repository LICENSE.txt)）。有公開程式碼不等於已被重現，也不代表目前版本與論文版本相同。