LEMON-Mapping
LEMON-Mapping 指出傳統多機位姿圖最佳化只把迴圈當作位姿節點間約束，忽略地圖幾何，導致重疊區發散與模糊。其框架包含：迴圈處理模組（剔除離群、分類迴圈並召回被誤刪的正確迴圈）；對多機地圖做空間 BA（孤立迴圈用 DBA、成群迴圈用 HBA）以消除重疊區不一致；再以兩步驟位姿圖最佳化，把 BA 精修後的局部精度傳遞到整張地圖。
本頁內容
LEMON-Mapping augments multi-session PGO with robust loop processing and spatial bundle adjustment in overlapping regions, then propagates BA-refined accuracy through a two-step PGO.
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: Velodyne (S3E), Avia and Ouster (GEODE), Avia (MARS-LVIG, R3LIVE), Mid360 (self-collected) (Tables I-II) |
|---|---|
| 原文測試平台 | not described in the paper: public multi-session datasets S3E, GEODE, MARS-LVIG and R3LIVE (platforms described only in the dataset papers)、self-collected Livox Mid360 dataset (Garage, Library, Yard, Laboratory, Flying Arena); carrier not stated |
| 狀態估計 | loop processing (outlier rejection, classification, recall) then spatial BA (DBA for isolated loops, HBA for clustered loops) and two-step pose-graph optimization |
| 資料關聯 | RING++ provides all intra- and inter-robot loop candidates; statistical outlier removal, GICP initialized by RING++ and RANSAC correspondence rejection with an inlier count and GICP fitness test verify loops; BFS region growing labels loops as clustered or isolated; rejected loops are recalled if their poses are within 2 m after the first PGO; spatial BA windows are spherical radius searches in a pose kd-tree around each loop; RING++ similarity (inter-robot) and a registration-based minimum-eigenvalue test (intra-robot) select sparse BA constraints for the last PGO (Sec. IV, V, VI-B). |
| 時間表示 | discrete poses |
| 去畸變 | 原文未報告 (inputs from LiDAR-inertial odometry) |
| 迴圈閉合 | robust loop outlier rejection with recall of wrongly removed loops |
| 全域最佳化 | window-based spatial BA in overlaps plus two-step PGO propagating local accuracy |
| 地圖表示 | multi-session point-cloud submaps |
| 先驗資訊 | multi-agent odometry and submaps; relative transforms between sessions unknown |
| 可輸出幾何 | merged globally consistent point-cloud map and trajectories |
| 計算需求 | Offline, C++ with ROS; hardware not reported. Runtime is governed by the number of loop closures (Dormitory, 62 loops, about 70 s); within LEMON-Mapping the spatial HBA stage for clustered loops takes about 67% to 80% and cluster preprocessing (PCA reordering and GICP) about 17% to 28% of runtime, isolated DBA and PGO each under 5%; peak memory stays bounded and below the HBA baseline, whereas a full BALM2 pipeline needs 30 GiB to 50 GiB even after aggregating f frames per submap (e.g., f = 30) (Sec. VII-H, Fig. 18). |
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 | Mid360歸入：Livox MID-360 | 方法輸入 | self-collected (Garage, Library, Yard, Laboratory, Flying Arena) | 原文未報告 | (Wang et al., 2026, Table I) |
| LiDAR | Velodyne (model not named) | 資料集感測器 | S3E | 原文未報告 | (Wang et al., 2026, Table II) |
| LiDAR | Avia歸入：Livox Avia | 資料集感測器 | GEODE (Inlandwaterways, Tunnelingtunnel); MARS-LVIG; R3LIVE | 原文未報告 | (Wang et al., 2026, Table II) |
| LiDAR | Ouster (model not named) | 資料集感測器 | GEODE (Stairs, Offroad) | 原文未報告 | (Wang et al., 2026, Table II) |
| LiDAR | DJI L1 LiDAR sensor | 參考或真值量測 | MARS-LVIG | high-precision point cloud processed with DJI Terra | (Wang et al., 2026, Sec. VII-D) |
作者報告的優勢與限制
優勢
- Merged all 10 multi-robot sequences of Table IV, whereas LAMM failed on 4 and DCL-SLAM on 6 (failure = ATE RMSE above 30 m) (Sec. VII-C, Table IV)
- Map quality against DJI L1 reference maps on MARS-LVIG Island, Town and Airport: AWD 0.25 m to 0.65 m and CD 0.43 m to 1.11 m, lower than LAMM on every metric (Sec. VII-D, Table V)
- S3E plane thickness 0.08 m and planarity 0.86 versus 0.12 m and 0.61 (DCL-SLAM) and 0.14 m and 0.59 (LAMM) (Table VI)
- Loop outlier rejection F1 higher than best-tuned PCM and GNC on three of four S3E sequences and comparable on Tunnel (Table VIII)
- Spatial BA gave the lowest z-drift and z-RMSE on all four self-collected single-robot scenes versus BALM2 and HBA (Table III)
- 100% merging success with 5, 10 and 20 sessions of R3LIVE data (Table X)
限制
- Information-matrix weights are heuristic; principled covariance estimation left to future work (Sec. VIII)
- Runtime grows with the number of loop closures; multi-threaded parallel windows left to future work (Sec. VII-H, VIII)
- Spatial BA without the last PGO can increase ATE because it breaks odometry continuity (Sec. VII-F, Table IX)
- Loop rejection precision and recall are computed against the framework's own optimized trajectory with a 5 m threshold, not independent labels (Sec. VII-E)
- Single-robot study has no ground truth; uses z-drift relative to the first frame and MME, and MME was not best on Garage or Laboratory (Sec. VII-B, Table III)
- Hardware for runtime and memory results not reported (Sec. VII-H)
營建工程相關證據
無工地資料。公開資料含 S3E Tunnel、GEODE Tunnelingtunnel 與 Stairs 等隧道與樓梯序列；自行蒐集資料（Mid360 光達）含車庫（室內）、圖書館、庭院、實驗室與飛行場（Sec. VII-A、Table I 至 II）。地圖品質以 DJI L1 光達經 DJI Terra 處理的高精度點雲作為 MARS-LVIG 參考地圖，AWD 0.25 至 0.65 m、CD 0.43 至 1.11 m（Sec. VII-D、Table V）；S3E 平面厚度 0.08 m（Table VI）。這些誤差等級遠大於工地尺寸檢核常用的公差（推論）。
原文驗證環境：公開基準、地下或隧道、獨立參考量測、跨場域
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 7 個比較組，合計 50 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 3 組列在最後，並連到性能比較頁。
Wang et al., 2026 · Table III 本方法 12 筆
表格設定（擷取紀錄原文）：Single-robot study on self-collected Mid360 data; spatial BA versus BALM2 (sliding window) and HBA, all on raw odometry without loop-based refinement; z-drift and z-RMSE relative to the z-value of the first frame; MME via MapEval (lower is better) (Wang et al., 2026, Table III)
MME，self-collected (Mid360) · Garage
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 Wang et al., 2026 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(Wang et al., 2026, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| HBA | -6.83 | (Wang et al., 2026, Table III) |
| BALM2 | -6.93 | (Wang et al., 2026, Table III) |
| Ours (spatial BA)本方法原文提出 | -6.86 | (Wang et al., 2026, Table III) |
Wang et al., 2026 · Table V 本方法 12 筆
表格設定（擷取紀錄原文）：Mapping quality against the MARS-LVIG ground-truth map (DJI L1 LiDAR processed with DJI Terra); AWD average Wasserstein distance, CD Chamfer distance, SCS spatial consistency score, MME mean map entropy; all lower is better; metrics as defined in MapEval (Wang et al., 2026, Table V)
AWD(m)，MARS-LVIG · Island
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 Wang et al., 2026 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(Wang et al., 2026, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LAMM | 1.3 m | (Wang et al., 2026, Table V) |
| Ours本方法原文提出 | 0.3 m | (Wang et al., 2026, Table V) |
Wang et al., 2026 · Table IV 本方法 10 筆
指標RMSE of the ATE (m)
表格設定（擷取紀錄原文）：Multi-robot localization; RMSE of ATE (m); failure (x) = any sequence with RMSE above 30 m; GEODE, MARS-LVIG and S3E sequences split into sessions with unknown relative transforms (Wang et al., 2026, Table IV)
RMSE of the ATE (m)，S3E · Campus 3
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 Wang et al., 2026 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(Wang et al., 2026, Table IV)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| DCL-SLAM | 17.89 m | (Wang et al., 2026, Table IV) |
| LAMM | 12.51 m | (Wang et al., 2026, Table IV) |
| Ours本方法原文提出 | 3.51 m | (Wang et al., 2026, Table IV) |
Wang et al., 2026 · Table IX 本方法 7 筆
指標RMSE of the ATE (m)
表格設定（擷取紀錄原文）：Ablation of the map merging module: first PGO only, first PGO plus spatial BA, and full LEMON-Mapping; RMSE of ATE (m) (Wang et al., 2026, Table IX)
RMSE of the ATE (m)，S3E · Campus 1
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 Wang et al., 2026 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(Wang et al., 2026, Table IX)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| FPGO | 9.97 m | (Wang et al., 2026, Table IX) |
| FPGO + BA | 10.05 m | (Wang et al., 2026, Table IX) |
| LEMON Full本方法原文提出 | 9.94 m | (Wang et al., 2026, Table IX) |
其他比較組
來源
Wang et al., 2026
（2026）LEMON-Mapping: Loop-Enhanced Large-Scale Multi-Session Point Cloud Merging and Optimization for Globally Consistent MappingIEEE Transactions on Automation Science and Engineering, 23:12318-12336
DOI 10.1109/tase.2026.3709653arXiv 2505.10018
同儕審查已出版已讀全文近十年
相關版本
- 預印本：arXiv 2505.10018 (v4, 2026-06-10) https://arxiv.org/abs/2505.10018