maplab 2.0
maplab 2.0 是以因子圖為核心的模組化、多模態建圖框架：一張地圖由多個任務（mission，即單次連續建圖時段）組成，頂點包含位姿、速度、IMU 偏差與地標，可整合視覺、光達與語意地標。新版加入 mapping server，把各機器人的子地圖先各自做局部最佳化與迴圈，再於全域層級做跨機迴圈與聯合最佳化；離線 console 提供批次最佳化、地圖合併、以 ICP/G-ICP 產生並以可切換約束抗離群的光達迴圈，以及 Voxblox 稠密重建。
本頁內容
maplab 2.0 is an open factor-graph mapping framework supporting multi-modal sensors, multi-session and multi-robot map merging with visual and LiDAR loop closures, and offline map processing.
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、IMU (optional; framework does not require an IMU)、multi-camera or monocular camera、GNSS (RTK, optional absolute constraints)、wheel encoders and RGB-D landmarks (supported interfaces, not evaluated) |
|---|---|
| 原文測試平台 | handheld、UAV (EuRoC MAV benchmark) |
| 狀態估計 | factor graph over vertices (6-DoF pose, velocity, IMU biases, landmarks) with batch optimization / bundle adjustment |
| 資料關聯 | visual landmarks from ORB detection with BRISK or FREAK binary descriptors (inverted multi-index matching), plus external float descriptors such as SuperPoint with SuperGlue tracking and SIFT with Lucas-Kanade tracking (PCA-compressed 256 to 32, FLANN matching); 2D-3D matches with covisibility filtering and P3P in RANSAC for visual loop closure, or landmark merging; 3D landmarks (RGB-D, LiDAR image keypoints) matched by 3D-3D RANSAC; LiDAR loop closures by ICP or G-ICP registration in the console |
| 時間表示 | discrete poses |
| 去畸變 | 原文未報告 (delegated to odometry source, e.g., FAST-LIO2) |
| 迴圈閉合 | visual and LiDAR intra- and inter-mission loop closures; loop edges as switchable constraints |
| 全域最佳化 | global multi-mission, multi-robot optimization in mapping server or offline console |
| 地圖表示 | factor-graph map of missions with attached sensor data; dense reconstruction via Voxblox plugin |
| 先驗資訊 | optional RTK GNSS absolute pose constraints |
| 可輸出幾何 | optimized poses, landmarks; dense volumetric reconstruction via Voxblox plugin; map data export (Sec. III-D) |
| 計算需求 | online mapping node and server plus offline console; Table II timings on Intel i7-8700 with Nvidia RTX 2080 GPU: 52 min of Hilti data processed in 98 to 267 min by the maplab 2.0 configurations (194 min for FAST-LIO2 + SP + B) versus 52 min for FAST-LIO2 alone; EuRoC server run 3 min 27 s versus 35 min 56 s sequentially (Sec. IV-B) |
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 | OS0-64歸入：Ouster OS0-64 | 資料集感測器 | HILTI 2021 SLAM Dataset | 原文未報告 | (Cramariuc et al., 2023, Sec. IV-A) |
| LiDAR | Ouster OS0-128 | 方法輸入 | 未標示 | 原文未報告 | (Cramariuc et al., 2023, Sec. IV-B) |
| 行動掃描設備 | handheld device with five cameras and an Ouster OS0-128 | 方法輸入 | 未標示 | 23 runs, more than two hours, about 10 km, indoor-outdoor transitions; OKVIS odometry | (Cramariuc et al., 2023, Sec. IV-B) |
| 慣性量測單元（IMU） | ADIS IMU | 資料集感測器 | HILTI 2021 SLAM Dataset | 原文未報告 | (Cramariuc et al., 2023, Sec. IV-A) |
| GNSS 接收器 | RTK GPS | 方法輸入 | 未標示 | optional absolute pose constraints where available | (Cramariuc et al., 2023, Sec. IV-B) |
| 相機 | five cameras (Hilti 2021 rig; model not stated in paper) | 資料集感測器 | HILTI 2021 SLAM Dataset | five cameras; frontal camera or stereo pair used for ROVIO/OKVIS odometry, all five for loop closure | (Cramariuc et al., 2023, Sec. IV-A) |
| 相機 | RGB-inertial sensor (cited as VersaVIS, ref. [52]) | 方法輸入 | custom indoor office semantic dataset | 原文未報告 | (Cramariuc et al., 2023, Sec. IV-D) |
| 運算硬體 | Intel i7-8700 | 執行運算平台 | 未標示 | 原文未報告 | (Cramariuc et al., 2023, Sec. IV-A) |
| 運算硬體 | Nvidia RTX 2080 GPU | 執行運算平台 | 未標示 | 原文未報告 | (Cramariuc et al., 2023, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Refines the best odometry: FAST-LIO2 + SP + B reduced Parking APE RMSE from 5.00 m (FAST-LIO2) to 0.21 m and matched or improved FAST-LIO2 on the other eight Hilti sequences (Table II)
- Among vision-based methods, maplab 2.0 configurations outperform ORB-SLAM3, RTAB-Map, maplab and ROVIO/OKVIS odometry on most sequences (Table II, Sec. IV-A)
- Handheld 23-run multi-session mapping over about two hours and 10 km in a training facility, merged with visual loops and optional RTK constraints (Sec. IV-B)
- Server-based multi-robot processing matched sequential multi-session accuracy on EuRoC (0.043 m average APE RMSE) while running about ten times faster (Sec. IV-B)
限制
- Loop-closure edges from LiDAR registration and visual matching use predefined, empirically chosen fixed covariances (Sec. III-A, III-D, IV-A); only the semantic-object loop-closure demo computes the constraint covariance (Sec. IV-D)
- (inference) LiDAR loop closure uses pairwise ICP/G-ICP; no BA-style point-cloud consistency refinement is described in the full text
- (inference) Every maplab 2.0 configuration needed more total time than the 52 min dataset duration: 98 to 267 min, and 194 min for the most accurate FAST-LIO2 + SP + B, on i7-8700 with RTX 2080 (Table II)
- LiDAR-image keypoint landmarks suffer outliers from missing points and moving objects (Sec. IV-C)
- On the two construction sequences the best maplab 2.0 result equals FAST-LIO2 alone (0.04 m and 0.07 m APE RMSE) (Table II); (inference) no construction-specific accuracy gain is shown
營建工程相關證據
使用 Hilti 2021 SLAM 資料集，作者描述其含室內辦公與戶外工地場景（Sec. IV-A）。Table II 中 Construction 1 與 Construction 2 的 APE RMSE，maplab 2.0（FAST-LIO2 + SP + B）為 0.04 m 與 0.07 m，與單獨 FAST-LIO2 相同，明顯改善只出現在 Parking 序列（5.00 m 降至 0.21 m）。大型訓練設施含倒塌建築與狹窄空間的 23 趟手持多時段建圖僅有定性展示（Sec. IV-B、Fig. 4）。
原文驗證環境：公開基準、施工中工地、獨立參考量測
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 2 個比較組，合計 44 筆紀錄。
Cramariuc et al., 2023 · Table II 本方法 40 筆
表格設定（擷取紀錄原文）：HILTI 2021 SLAM Dataset, RMSE of APE; seven baselines and four maplab 2.0 configurations (ROVIO + SIFT, OKVIS + SP + B, OKVIS + SP + B + ICP, FAST-LIO2 + SP + B); sensors used differ per method (icons in table); Total Time measured on Intel i7-8700 with Nvidia RTX 2080 for 52 min of data (Cramariuc et al., 2023, Table II)
RMSE of the absolute position error (APE)，HILTI 2021 SLAM Dataset · Construction 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 Cramariuc 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(Cramariuc et al., 2023, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB SLAM3 | 1.55 m | (Cramariuc et al., 2023, Table II) |
| LVI SAM | 0.13 m | (Cramariuc et al., 2023, Table II) |
| RTAB Map | 0.36 m | (Cramariuc et al., 2023, Table II) |
| maplab | 0.16 m | (Cramariuc et al., 2023, Table II) |
| ROVIO | 0.98 m | (Cramariuc et al., 2023, Table II) |
| OKVIS | 1.17 m | (Cramariuc et al., 2023, Table II) |
| FAST LIO2 | 0.04 m | (Cramariuc et al., 2023, Table II) |
| maplab 2.0: ROVIO + SIFT本方法原文提出 | 0.14 m | (Cramariuc et al., 2023, Table II) |
| maplab 2.0: OKVIS + SP + B本方法原文提出 | 0.08 m | (Cramariuc et al., 2023, Table II) |
| maplab 2.0: OKVIS + SP + B + ICP本方法原文提出 | 0.08 m | (Cramariuc et al., 2023, Table II) |
| maplab 2.0: FAST-LIO2 + SP + B本方法原文提出 | 0.04 m | (Cramariuc et al., 2023, Table II) |
Cramariuc et al., 2023 · Text Sec.IV-B 本方法 4 筆
資料集與序列EuRoC MAV · all 11 sequences
表格設定（擷取紀錄原文）：EuRoC 11 sequences with ROVIO and BRISK: parallel multi-robot mapping server vs sequential multi-session console workflow (Cramariuc et al., 2023, Text Sec.IV-B)
average RMSE APE，EuRoC MAV · all 11 sequences
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 Cramariuc 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(Cramariuc et al., 2023, Text Sec.IV-B)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| maplab 2.0 mapping server (parallel, 11 missions)本方法原文提出 | 0.043 m | (Cramariuc et al., 2023, Sec. IV-B) |
| maplab 2.0 sequential multi-session (mapping node + console)本方法原文提出 | 0.043 m | (Cramariuc et al., 2023, Sec. IV-B) |
來源
Cramariuc et al., 2023
（2023）maplab 2.0 – A Modular and Multi-Modal Mapping FrameworkIEEE Robotics and Automation Letters, 8(2):520-527
DOI 10.1109/lra.2022.3227865arXiv 2212.00654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：ethz-asl/maplab https://github.com/ethz-asl/maplab
程式碼：https://github.com/ethz-asl/maplab（授權：Apache-2.0 (LICENSE file)）。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。