PIN-SLAM
PIN-SLAM 以稀疏可最佳化的神經點編碼局部符號距離場（SDF），里程計採不需最近點配對的點對隱式 SDF 配準，並以局部地圖產生的描述子偵測迴圈、做位姿圖最佳化。因神經點隨所屬影格一起移動，迴圈修正後隱式地圖可保持全域一致並輸出網格。論文另在 Newer College 以毫米級 TLS 參考地圖評估網格精度，並在 Hilti-21（含營建工地序列）報告軌跡誤差。
本頁內容
LiDAR SLAM on elastic neural points encoding an SDF, with correspondence-free registration, loop closure and pose-graph-driven map deformation.
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 | Velodyne HDL64 (KITTI)、Ouster OS1-64 (MulRAN; Newer College long sequences; IPB-Car 2020)、OS1-128 (IPB-Car 2023)、OS0-128 (Newer College shorter sequences)、OS0-64 (Hilti-21, handheld)、32-beam LiDAR on a Spot robot (Nebula, qualitative)、synthetic RGB-D (Replica) |
|---|---|
| 原文測試平台 | vehicle、handheld、legged |
| 狀態估計 | correspondence-free scan-to-implicit-SDF registration with second-order (Levenberg-Marquardt) optimization and robust weights; pose graph optimization after loop closure |
| 資料關聯 | point-to-implicit SDF (no closest-point association) |
| 時間表示 | discrete poses |
| 去畸變 | constant-velocity prediction with per-point timestamps before odometry, then re-deskew with the odometry estimate (Sec. III-B); motion compensation disabled on KITTI because those scans are already deskewed (Sec. V-B1) |
| 迴圈閉合 | distance-based local loop check plus polar context descriptors computed from the local neural map; verification by scan-to-map registration |
| 全域最佳化 | pose graph optimization; neural points move with their associated frames (elastic map) |
| 地圖表示 | sparse optimizable neural points indexed by voxel hashing with a shared decoder to SDF |
| 先驗資訊 | none |
| 可輸出幾何 | mesh via marching cubes from the SDF; compact neural point map |
| 計算需求 | single NVIDIA Quadro A4000: full 7.1 Hz (0.14 s per frame), light 11.3 Hz (0.09 s per frame) average on KITTI 00-10 (Table XVI); frames with PGO occasionally >200 ms (Sec. V-F2); odometry and map optimization each take about 40% of runtime (Fig. 13); map memory 66.3 to 138.8 MB on KITTI 00, 05 and 08 and 76.8 MB on Newer College 02, about 0.3% to 1.1% of the raw point cloud (Table XV) |
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 | Velodyne HDL64歸入：Velodyne HDL-64E | 資料集感測器 | KITTI odometry | 64-beam, car-mounted; reference poses from GNSS-INS | (Pan et al., 2024, Sec. V-A1; Table I) |
| LiDAR | Ouster OS1-64 | 資料集感測器 | MulRAN | car-mounted; field of view partly blocked by the radar sensor; reference poses from GNSS-INS | (Pan et al., 2024, Sec. V-A1) |
| LiDAR | OS1-64歸入：Ouster OS1-64 | 資料集感測器 | Newer College | handheld, used for the two longer sequences | (Pan et al., 2024, Sec. V-A1) |
| LiDAR | OS0-128歸入：Ouster OS0-128 | 資料集感測器 | Newer College | handheld, used for the shorter sequences | (Pan et al., 2024, Sec. V-A1) |
| LiDAR | OS1-64 (2020) and OS1-128 (2023)歸入：Ouster OS1-64 | 資料集感測器 | IPB-Car | robot car in Bonn; self-collected | (Pan et al., 2024, Sec. V-A1) |
| LiDAR | OS0-64歸入：Ouster OS0-64 | 資料集感測器 | Hilti-21 | handheld; indoor offices, labs, basements and outdoor construction sites | (Pan et al., 2024, Sec. V-A1) |
| LiDAR | 32-beam LiDAR (model not named) | 資料集感測器 | Nebula | carried by a Spot1 robot moving back and forth in the Valentine Cave | (Pan et al., 2024, Fig. 7) |
| 地面雷射掃描儀（TLS） | geo-referenced terrestrial laser scanner (model not named) | 參考或真值量測 | IPB-Car | global map used for scan-to-map constraints with the OS1-128 in the factor graph that generates reference poses (fused with GNSS-INS, LiDAR odometry and loop closures) | (Pan et al., 2024, Sec. V-A1) |
| 地面雷射掃描儀（TLS） | survey-grade TLS point cloud map (scanner model not named) | 參考或真值量測 | Newer College | mm-level accuracy; reference poses obtained by aligning each scan to it; also the reference model for Table XI | (Pan et al., 2024, Sec. V-A1; Table XI caption) |
| 地面雷射掃描儀（TLS） | terrestrial laser scanner survey-grade map (model not named) | 參考或真值量測 | Nebula | used only for a qualitative mapping-error visualisation | (Pan et al., 2024, Fig. 7) |
| GNSS 接收器 | GNSS-INS (model not named) | 參考或真值量測 | KITTI odometry; MulRAN | poses regarded as the evaluation reference | (Pan et al., 2024, Sec. V-A1) |
| 全測站 | total station tracking system (model not named) | 參考或真值量測 | Hilti-21 | reference trajectories for some sequences (others from a motion capture system) | (Pan et al., 2024, Sec. V-A1) |
| 載具平台 | Spot1 robot | 資料集感測器 | Nebula | quadruped robot moving back and forth in a cave tunnel (Valentine Cave) | (Pan et al., 2024, Fig. 7; Sec. V-B4) |
| 運算硬體 | NVIDIA Quadro A4000 | 執行運算平台 | 未標示 | single GPU; 7.1 Hz full and 11.3 Hz light on KITTI | (Pan et al., 2024, Sec. V-F2; Table XVI) |
| 其他 | motion capture system (model not named) | 參考或真值量測 | Hilti-21 | reference trajectories for some sequences | (Pan et al., 2024, Sec. V-A1) |
作者報告的優勢與限制
優勢
- Localization on par or better than LiDAR odometry/SLAM baselines (abstract; Sec. V-B)
- Best completeness, Chamfer and F-score among compared mesh-producing methods on two Newer College scenes vs TLS reference (Sec. V-D1; Table XI)
- Sensor-rate operation on a moderate GPU (Sec. V-F2)
限制
- No IMU; future work (Sec. VI)
- Fixed neural point resolution (Sec. VI)
- Loop/PGO frames exceed 200 ms (Sec. V-F2)
- Hilti-21 sequences are short without explicit loops, so odometry and SLAM are not distinguished there (Sec. V-B2)
- Loop closure detection recall at Top-1 averages 94.5%, below BEVPlace at 99.3% (Table X)
- (derived from Table VI) on the loop-free IPB-Car 2023-0 drive PIN-SLAM has 87.59 m ATE RMSE, equal to PIN-LO, and all compared methods exceed 78 m
營建工程相關證據
作者在 Hilti-21 資料集評估軌跡；論文 Sec. V-A1 說明該資料集含戶外營建工地序列，參考軌跡由全測站追蹤系統或動作捕捉系統量測。Table VIII 的 cons2 欄對應營建工地序列（名稱對應為推論）；PIN-SLAM 在 cons2 的 ATE RMSE 為 0.41 m，為所比較方法中最佳，但也是其六個 Hilti-21 序列中最大的誤差；作者並指出這些序列較短且無明確迴圈。此為公開基準的軌跡層級證據，未評估工地點雲或網格品質。另在 Nebula 天然洞穴資料以 TLS 地圖做定性誤差展示（非營建隧道）。
原文驗證環境：公開基準、獨立參考量測、施工中工地、地下或隧道
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 15 個比較組，合計 99 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 11 組列在最後，並連到性能比較頁。
Pan et al., 2025 · Table II 本方法 16 筆
表格設定（擷取紀錄原文）：Oxford Spires surface reconstruction against the millimetre-accurate Leica RTC360 TLS reference map; localization disabled and ground-truth poses used for all methods; OpenMVS and Nerfacto results taken from the benchmark (offline batch); meshes at 0.1 m resolution; F-score threshold 0.1 m; precision and recall columns not extracted (Pan et al., 2025, Table II)
Accuracy error，Oxford Spires · Blenheim Palace 05
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 Pan 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(Pan et al., 2025, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| OpenMVS [5] (offline) | 0.126 m | (Pan et al., 2025, Table II) |
| Nerfacto [62] (offline) | 0.302 m | (Pan et al., 2025, Table II) |
| GSS [11] | 0.204 m | (Pan et al., 2025, Table II) |
| VDB-Fusion [67] | 0.098 m | (Pan et al., 2025, Table II) |
| PIN-SLAM [51]本方法 | 0.078 m | (Pan et al., 2025, Table II) |
| PINGS (Ours)原文提出 | 0.072 m | (Pan et al., 2025, Table II) |
Pan et al., 2024 · Table VII 本方法 16 筆
指標ATE RMSE [m]
表格設定（擷取紀錄原文）：Newer College handheld LiDAR (OS1-64 long sequences, OS0-128 shorter sequences), reference poses from registering each scan to a survey-grade TLS map; all sequences contain loops; failure marked by a cross, '-' not reported or unavailable; ATE RMSE [m] with Umeyama trajectory alignment (Sec. V-B2; whether scale is estimated is not stated) (Pan et al., 2024, Table VII)
ATE RMSE [m]，Newer College · 01
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。
- 未報告（沒有數值，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 Pan et al., 2024 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(Pan et al., 2024, Table VII)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| F-LOAM [82] | 6.74 m | (Pan et al., 2024, Table VII) |
| KISS-ICP [78] | 0.62 m | (Pan et al., 2024, Table VII) |
| SuMa [4] | 2.03 m | (Pan et al., 2024, Table VII) |
| MULLS [52] | 2.51 m | (Pan et al., 2024, Table VII) |
| MD-SLAM [13] | 無數值未報告註記（擷取紀錄）：原文未報告 ('-') | (Pan et al., 2024, Table VII) |
| SC-LeGO-LOAM [68, 27] | 無數值未報告註記（擷取紀錄）：原文未報告 ('-') | (Pan et al., 2024, Table VII) |
| PIN-LO本方法原文提出 | 2.08 m | (Pan et al., 2024, Table VII) |
| PIN-SLAM本方法原文提出 | 0.43 m | (Pan et al., 2024, Table VII) |
Guadagnino et al., 2025a · Table III 本方法 9 筆
指標ATE [m] (evo)
表格設定（擷取紀錄原文）：ATE from evo; the paired relative KITTI metric (%) is omitted; '-' means the run failed because errors exceeded a sequence-specific threshold (version of record); same KISS-SLAM configuration for all runs; values averaged over three runs per scene (Guadagnino et al., 2025a, Table III)
ATE [m] (evo)，HeLiPR · Bridge Aeva
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。
- 失敗
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 Guadagnino et al., 2025a 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(Guadagnino et al., 2025a, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| PIN-SLAM本方法 | 無數值失敗註記（擷取紀錄）：failed ('-': error exceeded a sequence-specific threshold) | (Guadagnino et al., 2025a, Table III) |
| SuMa | 無數值失敗註記（擷取紀錄）：failed ('-': error exceeded a sequence-specific threshold) | (Guadagnino et al., 2025a, Table III) |
| CT-ICP | 無數值失敗註記（擷取紀錄）：failed ('-': error exceeded a sequence-specific threshold) | (Guadagnino et al., 2025a, Table III) |
| MULLS | 356.06 m | (Guadagnino et al., 2025a, Table III) |
| Ours (KISS-SLAM)原文提出 | 98.61 m | (Guadagnino et al., 2025a, Table III) |
Pan et al., 2025 · Table III 本方法 8 筆
表格設定（擷取紀錄原文）：In-house car dataset (Bonn), full sequences; reference poses from offline LiDAR bundle adjustment with RTK-GNSS, point cloud alignment and geo-referenced TLS constraints; odometry methods above, SLAM methods below; ATE alignment not stated (Pan et al., 2025, Table III)
ARTE [%] (average relative translation error)，in-house car dataset · Seq. 1 (5.0 km)
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 Pan 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(Pan et al., 2025, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| F-LOAM [69] | 1.96% | (Pan et al., 2025, Table III) |
| KISS-ICP [68] | 1.49% | (Pan et al., 2025, Table III) |
| PIN odometry [51]本方法 | 0.95% | (Pan et al., 2025, Table III) |
| PINGS odometry原文提出 | 0.73% | (Pan et al., 2025, Table III) |
| SuMa [3] | 5.55% | (Pan et al., 2025, Table III) |
| MULLS [50] | 2.23% | (Pan et al., 2025, Table III) |
| PIN-SLAM [51]本方法 | 1% | (Pan et al., 2025, Table III) |
| PINGS (Ours)原文提出 | 0.68% | (Pan et al., 2025, Table III) |
其他比較組
列出其餘 11 個比較組
- Pan et al., 2024 · Table XI
- Guadagnino et al., 2025a · Table IV
- Guadagnino et al., 2025a · Table V
- Pan et al., 2024 · Table VIII
- Pan et al., 2024 · Table XVI
- Guadagnino et al., 2025a · Table II
- Pan et al., 2024 · Table IV
- Pan et al., 2024 · Table IX
- Tosi et al., 2026 · Table XI
- Ghadimzadeh Alamdari et al., 2025 · Table 3
- Guadagnino et al., 2025a · Table I
來源
Pan et al., 2024
（2024）PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map ConsistencyIEEE Transactions on Robotics, 40, 4045-4064
DOI 10.1109/tro.2024.3422055arXiv 2401.09101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
- 預印本：arXiv:2401.09101 https://arxiv.org/abs/2401.09101
- 程式碼釋出：PRBonn/PIN_SLAM https://github.com/PRBonn/PIN_SLAM
程式碼：https://github.com/PRBonn/PIN_SLAM（授權：MIT）。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。