ORB-SLAM3
ORB-SLAM3 在 ORB-SLAM2 基礎上加入緊耦合的視覺慣性（visual-inertial）最大後驗估計，包括 IMU 初始化階段，並支援針孔與魚眼相機。其 Atlas 多地圖機制在追蹤失敗時另起新地圖，重訪時再以改良召回率的場所辨識（place recognition）將地圖合併，使 BA 可使用時間上相隔很遠甚至跨作業階段的共視關鍵影格。系統輸出仍是稀疏地圖點與關鍵影格軌跡。
本頁內容
ORB-SLAM3 adds tightly integrated MAP visual-inertial estimation (including IMU initialisation), fisheye support and an Atlas multi-map with map merging, reusing co-visible keyframes across time and sessions in BA.
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 (pin-hole or fisheye; rectification not required)、RGB-D (supported by the library; no RGB-D experiment reported)、IMU |
|---|---|
| 原文測試平台 | UAV、handheld |
| 狀態估計 | Keyframe-based MAP estimation: visual or visual-inertial BA with IMU preintegration on manifold and Huber-robust reprojection terms; tracking optimises only the states of the last two frames with map points fixed; local mapping optimises a sliding window of keyframes and their points with covisible keyframes fixed. IMU initialisation in three MAP steps: 2 s of monocular visual-only BA (10 keyframes at 4 Hz), inertial-only MAP for scale, gravity direction, biases and velocities with a bias prior, then joint visual-inertial MAP; visual-inertial BA again 5 and 15 s after initialisation (scale error about 5% after 2 s and 1% after 15 s) |
| 資料關聯 | ORB features and reprojection; DBoW2 keyframe database with a new place-recognition method of improved recall (Sec. III; abstract) |
| 時間表示 | discrete keyframe states; IMU residuals between frames in visual-inertial mode (Sec. III) |
| 去畸變 | 不適用 |
| 迴圈閉合 | For each new keyframe, DBoW2 returns the three most similar Atlas keyframes not covisible with it; for each candidate a local window (candidate plus best covisible keyframes) is aligned by RANSAC with Horn's method on 3D-3D matches (Sim(3) for monocular or immature monocular-inertial maps, SE(3) otherwise), refined by guided matching and bidirectional reprojection optimisation, and verified in three covisible keyframes already in the map instead of three consecutive BoW detections; mature visual-inertial maps also require pitch and roll below a threshold. A match in the active map triggers loop correction, a match in another map triggers map merging |
| 全域最佳化 | Loop: welding window with point fusion, essential-graph pose-graph optimisation, then global BA in an independent thread; in the visual-inertial case global BA runs only if the number of keyframes is below a threshold. Map merge: welding BA over the merge window (stored-map keyframes outside it fixed) followed by essential-graph optimisation of the whole merged map with the welding area fixed. IMU scale and gravity refinement every 10 s until 100 keyframes or 75 s after initialisation |
| 地圖表示 | Atlas of disconnected sparse maps (map points plus keyframes), one active (Sec. III) |
| 先驗資訊 | No prior map required; maps of earlier sessions stored in the Atlas can be reused and merged. Requires camera intrinsics, body-to-camera extrinsics T_CB from calibration and, for stereo, a constant relative SE(3) between the cameras; IMU initialisation uses a prior that keeps biases near zero |
| 可輸出幾何 | Keyframe trajectories and sparse map points (on EuRoC V202 the four ORB-SLAM3 configurations keep 9,686 to 14,245 map points and 135 to 332 keyframes, Table VI); no dense reconstruction is produced or evaluated in the paper |
| 計算需求 | Intel Core i7-7700 CPU at 3.6 GHz with 32 GB memory, CPU only; real time at 30 to 40 frames and 3 to 6 keyframes per second. On EuRoC V202 the mean tracking time is 21.52 ms (monocular), 31.48 ms (stereo), 23.22 ms (monocular-inertial) and 33.05 ms (stereo-inertial) against 37.87 ms for ORB-SLAM2 stereo (Table VI). In the V201 to V203 multi-session run (Table VII), place recognition totals 3.45 to 5.89 ms per keyframe, map merging 120.63 to 287.33 ms, and the loop full BA up to 4134.94 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 |
|---|---|---|---|---|---|
| 慣性量測單元（IMU） | EuRoC IMU (model not stated) | 資料集感測器 | EuRoC | 200 Hz | (Campos et al., 2021, Table VI) |
| 慣性量測單元（IMU） | TUM-VI rig IMU (model not stated) | 資料集感測器 | TUM-VI | part of the hand-held stereo-inertial rig | (Campos et al., 2021, Sec. VII-B) |
| 雙目相機 | EuRoC stereo camera (model not stated) | 資料集感測器 | EuRoC | 752x480 at 20 Hz; 1000 to 1200 ORB features per image in the experiments | (Campos et al., 2021, Table VI) |
| 雙目相機 | TUM-VI hand-held fisheye stereo-inertial rig, cameras (model not stated) | 資料集感測器 | TUM-VI | fisheye stereo; CLAHE equalisation applied; 1500 ORB points per image (monocular-inertial) or 1000 (stereo-inertial) | (Campos et al., 2021, Sec. VII-B) |
| 載具平台 | drone (EuRoC micro aerial vehicle) | 資料集感測器 | EuRoC | 11 sequences in a machine hall and two Vicon rooms | (Campos et al., 2021, Abstract; Sec. VII-C) |
| 運算硬體 | Intel Core i7-7700 | 執行運算平台 | 未標示 | 3.6 GHz, 32 GB memory, CPU only | (Campos et al., 2021, Sec. VII) |
作者報告的優勢與限制
優勢
- Stereo-inertial average accuracy 3.5 cm on EuRoC and 9 mm under quick hand-held motion in TUM-VI room sequences (abstract)
- Survives long periods of poor visual information by starting and later merging maps (abstract)
限制
- Main failure case is low-texture environments (Sec. VIII)
- IMU is difficult to initialise with slow motion or without roll/pitch excitation, e.g. a car on flat ground (Sec. VIII)
- Monocular-inertial cannot estimate depth under pure rotation during exploration (Sec. VIII)
- Long outdoor TUM-VI sequences: scarce close features let inertial parameters (scale, accelerometer bias) drift, giving errors of 10 to 70 m; points beyond 20 m were discarded outdoors to suppress sky features (Sec. VII-B)
- Monocular ORB-SLAM3 did not complete V203 in single-session EuRoC (Table II)
- Stereo-inertial was less accurate than monocular-inertial in two Machine Hall sequences, attributed to greater scene depth (Sec. VII-A)
- In some dark TUM-VI slide sequences VINS-Mono and BASALT, which track with Lucas-Kanade, were more accurate (Sec. VII-B)
- A visual-inertial map lost within 15 s of IMU initialisation is discarded (Sec. V-D)
營建工程相關證據
論文未報告營建工地測試；評估資料為 EuRoC（含工業廠房）與 TUM-VI。作者指出低紋理環境為主要失效情境，與施工中大面積素面牆體的條件相關，但本文尚未找到直接工地驗證（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 42 個比較組，合計 294 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 38 組列在最後，並連到性能比較頁。
Campos et al., 2021 · Table II 本方法 48 筆
表格設定（擷取紀錄原文）：EuRoC single session, RMS ATE (m); ORB-SLAM3 median of 10 executions, Sim(3) alignment for monocular and SE(3) otherwise; other systems as reported by their authors except VINS-Mono and VINS-Fusion (run by the ORB-SLAM3 authors with default settings) and MCSKF, OKVIS, ROVIO (values from ref. [78]); ORB-SLAM, ORBSLAM-VI and BASALT use keyframe trajectories, ORB-SLAM and ORBSLAM-VI raw ground truth. Row cap: per-sequence values kept only for ORB-SLAM3, ORB-SLAM2, VINS-Mono and OKVIS; averages kept for all systems (Campos et al., 2021, Table II)
RMS ATE (m)，EuRoC · MH01
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 Campos 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(Campos et al., 2021, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM3 (stereo)本方法原文提出 | 0.029 m | (Campos et al., 2021, Table II) |
| ORB-SLAM3 (monocular-inertial)本方法原文提出 | 0.062 m | (Campos et al., 2021, Table II) |
| ORB-SLAM3 (stereo-inertial)本方法原文提出 | 0.036 m | (Campos et al., 2021, Table II) |
Campos et al., 2021 · Table IV 本方法 28 筆
表格設定（擷取紀錄原文）：TUM-VI room sequences (ground truth over the whole trajectory), RMS ATE (m) of ORB-SLAM3 in four sensor configurations, median of 3 executions; monocular aligned with 7 DoF, the others with 6 DoF (Campos et al., 2021, Table IV)
RMS ATE (m)，TUM-VI · room1
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 Campos 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(Campos et al., 2021, Table IV)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM3 (stereo)本方法原文提出 | 0.077 m | (Campos et al., 2021, Table IV) |
| ORB-SLAM3 (monocular-inertial)本方法原文提出 | 0.009 m | (Campos et al., 2021, Table IV) |
| ORB-SLAM3 (stereo-inertial)本方法原文提出 | 0.008 m | (Campos et al., 2021, Table IV) |
Xiao et al., 2025 · Table II 本方法 18 筆
表格設定（擷取紀錄原文）：Tracking accuracy with rpg trajectory evaluation: t_rel = average translational RMSE drift (%), r_rel = average rotational RMSE drift (deg/100 m), t_abs = ATE RMSE (m); reference trajectories from R3LIVE (not an independent measurement); alignment not stated; IMU not used by LiV-GS; '-' entries reported without explanation (text says indoor-oriented 3DGS SLAM methods degrade or fail on some outdoor sequences) (Xiao et al., 2025, Table II)
t_rel (average translational RMSE drift)，NTU4DRadLM · cp
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 Xiao 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(Xiao et al., 2025, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| NeRF-LOAM | 2.943% | (Xiao et al., 2025, Table II) |
| HDL-graph-SLAM | 1.264% | (Xiao et al., 2025, Table II) |
| ORB-SLAM3本方法 | 1.356% | (Xiao et al., 2025, Table II) |
| SplaTAM | 無數值未報告註記（擷取紀錄）：not reported ('-' in table) | (Xiao et al., 2025, Table II) |
| MonoGS | 4.171% | (Xiao et al., 2025, Table II) |
| Gaussian-SLAM | 1.249% | (Xiao et al., 2025, Table II) |
| GS-ICP-SLAM | 5.471% | (Xiao et al., 2025, Table II) |
| Ours原文提出 | 0.234% | (Xiao et al., 2025, Table II) |
Campos et al., 2021 · Table V 本方法 16 筆
表格設定（擷取紀錄原文）：EuRoC multi-session: all sessions of one environment processed sequentially, single global alignment, RMS ATE (m); ORB-SLAM3 median of 5 executions against processed ground truth; CCM-SLAM and VINS values as reported by their authors; '-' cells and scale-error rows omitted (Campos et al., 2021, Table V)
Multi-session RMS ATE (m)，EuRoC · MH01-03
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 Campos 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(Campos et al., 2021, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM3 (stereo)本方法原文提出 | 0.028 m | (Campos et al., 2021, Table V) |
| ORB-SLAM3 (monocular-inertial)本方法原文提出 | 0.037 m | (Campos et al., 2021, Table V) |
| ORB-SLAM3 (stereo-inertial)本方法原文提出 | 0.041 m | (Campos et al., 2021, Table V) |
其他比較組
列出其餘 38 個比較組
- Lipson et al., 2024 · Table 2b
- Teed & Deng, 2021 · Table 3
- Teed et al., 2023 · Table 3
- Teed & Deng, 2021 · Table 4
- Cramariuc et al., 2023 · Table II
- Zhang et al., 2023b · Table 3
- Nguyen et al., 2022b · Table IV
- Zhao et al., 2024b · Table 4
- Lee et al., 2025b · Table 7
- Huang et al., 2024c · Table 1
- Yan et al., 2026a · Table 1
- Huang et al., 2024c · Table 2
- Nguyen et al., 2022b · Table II
- Nguyen et al., 2022b · Table III
- Zhao et al., 2024b · Table 2
- Deng & Gan, 2026 · Table 11
- Matsuki et al., 2024 · Table 14
- Ha et al., 2024 · Table 2
- Campos et al., 2021 · Table VI
- Huang et al., 2024c · Table 3
- Yan et al., 2026a · Table 2
- Deng & Gan, 2026 · Table 1
- Xiao et al., 2025 · Table IV
- Boche et al., 2025 · Text Sec.VI-H
- Keetha et al., 2024 · Table 1
- Murai et al., 2025 · Fig. 5 table
- Murai et al., 2025 · Table 1
- Boche et al., 2025 · Table II
- Boche et al., 2025 · Table VI
- Deng & Gan, 2026 · Table 2
- Teed et al., 2023 · Table 1
- Lipson et al., 2024 · Table 1
- Lipson et al., 2024 · Table 3
- Teed & Deng, 2021 · Table 5
- Ha et al., 2024 · Table 1
- Lee et al., 2025b · Table 8 (Total column)
- Boche et al., 2025 · Table IV
- Maggio et al., 2025 · Table 2
來源
Campos et al., 2021
（2021）ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual–Inertial, and Multimap SLAMIEEE Transactions on Robotics, 37(6):1874-1890
DOI 10.1109/tro.2021.3075644arXiv 2007.11898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-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM (arXiv) https://arxiv.org/abs/2007.11898
- 程式碼釋出：ORB_SLAM3 https://github.com/UZ-SLAMLab/ORB_SLAM3
程式碼：https://github.com/UZ-SLAMLab/ORB_SLAM3（授權：GPLv3 (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。