ORB-SLAM2
ORB-SLAM2 將 ORB-SLAM 擴充到雙目（stereo）與 RGB-D 相機，把近距與遠距雙目特徵納入 BA，使尺度可觀測，迴圈閉合改以剛體 SE(3) 位姿圖最佳化並在另一執行緒進行全域 BA。系統另提供只做定位的地圖重用模式。作者明言目標是長期且全域一致的定位，而非最精細的稠密重建；論文中的稠密點雲是以估計的關鍵影格位姿反投影感測器深度圖所得。
本頁內容
ORB-SLAM2 extends ORB-SLAM to stereo and RGB-D with metric-scale BA, SE(3) loop closure plus full BA, and a map-reuse localization mode; its dense clouds are depth maps back-projected from estimated keyframe poses, not a jointly optimised dense map.
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 | monocular camera、stereo、RGB-D |
|---|---|
| 原文測試平台 | handheld、UAV、vehicle |
| 狀態估計 | BA with monocular and stereo constraints (motion-only, local, and full BA in a separate thread after pose-graph optimisation) (Sec. III, III-D) |
| 資料關聯 | ORB extracted on both rectified stereo images (or on the RGB image); a stereo keypoint (uL, vL, uR) comes from matching each left ORB along the same row with subpixel patch-correlation refinement; for RGB-D the depth d is converted to a virtual right coordinate uR = uL - fx*b/d with b approximated to 8 cm for Kinect and Asus Xtion; keypoints with depth below 40 times the baseline are close (triangulated from one frame, carry scale), others far (triangulated only from multiple views), unmatched ones stay monocular; DBoW2 place recognition as in ORB-SLAM |
| 時間表示 | discrete poses (keyframes) |
| 去畸變 | 不適用 |
| 迴圈閉合 | DBoW2 detection with geometric validation; rigid-body SE(3) pose-graph when stereo/depth makes scale observable, followed by full BA (Sec. III-D) |
| 全域最佳化 | pose-graph optimisation then full BA in a separate thread, with corrections propagated through the spanning tree (Sec. III-D) |
| 地圖表示 | sparse map points plus keyframes (covisibility graph, spanning tree) |
| 先驗資訊 | No prior map in SLAM mode; the Localization Mode reuses a previously built map with local mapping and loop closing disabled. Assumes rectified stereo with known focal length, principal point and baseline; the RGB-D structured-light baseline is approximated to 8 cm; the 4% depth scale bias of TUM freiburg2 sequences was compensated in the authors' runs |
| 可輸出幾何 | keyframe trajectory and sparse map points; the dense point clouds shown are obtained by back-projecting sensor depth maps from estimated keyframe poses (Sec. IV-C, Fig. 7) |
| 計算需求 | Intel Core i7-4790 desktop with 16 GB RAM, CPU only; each sequence run 5 times (median accuracy reported). Mean tracking time per frame 41.66 ms on EuRoC V2_02 (stereo 752x480, 20 Hz), 49.47 ms on KITTI 07 (stereo 1226x370, 10 Hz) and 25.58 ms on TUM fr3_office (RGB-D 640x480, 30 Hz), each below the frame period; local mapping 129.52 to 267.33 ms per keyframe; full BA after the single loop 349.25 to 1640.96 ms in a separate thread |
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 |
|---|---|---|---|---|---|
| 雙目相機 | KITTI stereo camera (model not stated) | 資料集感測器 | KITTI odometry | baseline about 54 cm, 10 Hz, 1240x376 after rectification | (Mur-Artal & Tardos, 2017, Sec. IV-A) |
| 雙目相機 | EuRoC stereo camera (model not stated) | 資料集感測器 | EuRoC | baseline about 11 cm, WVGA images at 20 Hz | (Mur-Artal & Tardos, 2017, Sec. IV-B) |
| RGB-D 相機 | Kinect | 方法輸入 | 未標示 | structured-light projector to infrared camera baseline approximated to 8 cm | (Mur-Artal & Tardos, 2017, Sec. III-A) |
| RGB-D 相機 | Asus Xtion | 方法輸入 | 未標示 | structured-light projector to infrared camera baseline approximated to 8 cm | (Mur-Artal & Tardos, 2017, Sec. III-A) |
| RGB-D 相機 | TUM RGB-D sensor (model not stated) | 資料集感測器 | TUM RGB-D | 640x480 at 30 Hz; freiburg2 depth maps with about 4% scale bias | (Mur-Artal & Tardos, 2017, Sec. IV-C, Table IV) |
| 載具平台 | car (KITTI) | 資料集感測器 | KITTI odometry | urban and highway driving | (Mur-Artal & Tardos, 2017, Sec. IV-A) |
| 載具平台 | micro aerial vehicle (EuRoC) | 資料集感測器 | EuRoC | flights in two rooms and a large industrial environment | (Mur-Artal & Tardos, 2017, Sec. IV-B) |
| 運算硬體 | Intel Core i7-4790 | 執行運算平台 | 未標示 | desktop computer, 16 GB RAM | (Mur-Artal & Tardos, 2017, Sec. IV) |
作者報告的優勢與限制
優勢
- Evaluation on 29 public sequences (abstract)
- Real-time on CPU (abstract)
- Zero-drift localization in already mapped areas (Sec. V)
限制
- Goal is globally consistent localization rather than the most detailed dense reconstruction (Sec. II-B)
- Stereo run failed on EuRoC V2_03_difficult (marked X in Table II)
- freiburg2 TUM depth maps had a 4% scale bias that authors compensated, which may partly explain their better results (Sec. IV-C)
- Stereo tracking is lost in parts of EuRoC V2_03_difficult because of severe motion blur; the authors note the sequence can be processed with IMU information (Sec. IV-B)
- The KITTI 09 loop, visible only in a few frames at the end, is not detected (Sec. IV-A)
- KITTI 01 highway: translation error is worse because few close points can be tracked at high speed and 10 Hz (Sec. IV-A, Table I)
- Localization mode assumes no significant environment changes; its visual-odometry matches can accumulate drift in unmapped regions (Sec. III-F)
營建工程相關證據
論文未於營建工地測試；EuRoC 包含工業廠房（machine hall）無人機序列，但不等同營建工地。稠密點雲僅為深度圖依位姿反投影的視覺化，其幾何誤差未經評估（推論）。
原文驗證環境：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 33 個比較組，合計 233 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 29 組列在最後，並連到性能比較頁。
Mur-Artal & Tardos, 2017 · Table I 本方法 33 筆
表格設定（擷取紀錄原文）：KITTI odometry training sequences, stereo: average relative translation error trel (%) and rotation error rrel (deg/100 m) per the KITTI metric, and absolute translation RMSE tabs (m); ORB-SLAM2 median of 5 runs; Stereo LSD-SLAM values as published by its authors (Mur-Artal & Tardos, 2017, Table I)
trel (%)，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 Mur-Artal & Tardos, 2017 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(Mur-Artal & Tardos, 2017, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM2 (stereo)本方法原文提出 | 0.7% | (Mur-Artal & Tardos, 2017, Table I) |
| Stereo LSD-SLAM | 0.63% | (Mur-Artal & Tardos, 2017, Table I) |
Yunus et al., 2021 · Table I 本方法 18 筆
指標ATE RMSE (m)
表格設定（擷取紀錄原文）：Translation ATE RMSE (m); ORB-SLAM2 and SP-SLAM run without bundle adjustment and loop closure for fairness; 'x' tracking failure, '-' result not available; the number of frames using Manhattan-frame tracking is also listed in the table (not extracted) (Yunus et al., 2021, Table I)
ATE RMSE (m)，ICL-NUIM · lr-kt0
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 Yunus 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(Yunus et al., 2021, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Ours (ManhattanSLAM)原文提出 | 0.007 m | (Yunus et al., 2021, Table I) |
| S-SLAM [11] | 無數值未報告註記（擷取紀錄）：result not available | (Yunus et al., 2021, Table I) |
| RGBD-SLAM [12] (Li et al. 2020) | 0.006 m | (Yunus et al., 2021, Table I) |
| ORB-SLAM2 [6] (BA and loop closure disabled)本方法 | 0.014 m | (Yunus et al., 2021, Table I) |
| SP-SLAM [5] (BA and loop closure disabled) | 0.019 m | (Yunus et al., 2021, Table I) |
| L-SLAM [10] | 0.015 m | (Yunus et al., 2021, Table I) |
Lipson et al., 2024 · Table 2b 本方法 13 筆
表格設定（擷取紀錄原文）：KITTI odometry sequences 00-10, monocular ATE; X = failure, '-' = average not computed; values checked against the ECCV 2024 version of record (same table numbering) (Lipson et al., 2024, Table 2b)
ATE[m]，KITTI · seq 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 Lipson 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(Lipson et al., 2024, Table 2b)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM2 [ 18 ]本方法 | 8.27 m | (Lipson et al., 2024, Table 2(b)) |
| ORB-SLAM3 [ 2 ] | 6.77 m | (Lipson et al., 2024, Table 2(b)) |
| LDSO [ 11 ] | 9.32 m | (Lipson et al., 2024, Table 2(b)) |
| DROID-VO [ 31 ] | 98.43 m | (Lipson et al., 2024, Table 2(b)) |
| DPVO [ 32 ] | 113.21 m | (Lipson et al., 2024, Table 2(b)) |
| DROID-SLAM [ 31 ] | 92.1 m | (Lipson et al., 2024, Table 2(b)) |
| DPV-SLAM原文提出 | 112.8 m | (Lipson et al., 2024, Table 2(b)) |
| DPV-SLAM++原文提出 | 8.3 m | (Lipson et al., 2024, Table 2(b)) |
Mur-Artal & Tardos, 2017 · Table IV 本方法 12 筆
表格設定（擷取紀錄原文）：Mean time per thread task (ms, mean of the thread total); loop and full-BA values are single measurements because each sequence has one loop; component times not transcribed (Mur-Artal & Tardos, 2017, Table IV)
Tracking total per frame (mean)，EuRoC · V2_02
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Mur-Artal & Tardos, 2017 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM2本方法原文提出硬體：Intel Core i7-4790, 16 GB RAM | 41.66 ms | (Mur-Artal & Tardos, 2017, Table IV) |
其他比較組
列出其餘 29 個比較組
- Campos et al., 2021 · Table II
- Rogers et al., 2020 · Table I
- Gao et al., 2018 · Table I
- Mur-Artal & Tardos, 2017 · Table II
- Teed & Deng, 2021 · Table 4
- Yan et al., 2026b · Table III
- Schöps et al., 2019 · Table 3
- Matsuki et al., 2024 · Table 1
- Yang et al., 2020a · Table 4
- Mur-Artal & Tardos, 2017 · Table III
- Zhu et al., 2021 · Table II
- Liso et al., 2024 · Table 2
- Sandström et al., 2023 · Table 3
- Keetha et al., 2024 · Table 1
- Chen et al., 2025a · Table 7
- Yan et al., 2024 · Table 2
- Peng et al., 2024 · Table 2
- Schöps et al., 2019 · Table 2
- Wang et al., 2023a · Table 4
- Sucar et al., 2021 · Table 3
- Zhu et al., 2022a · Table 2
- Yan et al., 2026b · Table VI
- Teed & Deng, 2021 · Fig. 4 table
- Yunus et al., 2021 · Table II
- Sumikura et al., 2019 · Fig. 4 table
- Sumikura et al., 2019 · Fig. 6 table
- Lipson et al., 2024 · Table 1
- Teed & Deng, 2021 · Table 5
- Campos et al., 2021 · Table VI
來源
Mur-Artal & Tardos, 2017
（2017）ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D CamerasIEEE Transactions on Robotics, 33(5):1255-1262
DOI 10.1109/tro.2017.2705103arXiv 1610.06475程式碼
同儕審查已出版已讀全文近十年
相關版本
- 預印本：ORB-SLAM2: an Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras (arXiv) https://arxiv.org/abs/1610.06475
- 程式碼釋出：ORB_SLAM2 https://github.com/raulmur/ORB_SLAM2
程式碼：https://github.com/raulmur/ORB_SLAM2（授權：GPLv3 (LICENSE.txt header)）。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。