SLAM 點雲圖鑑Point clouds for construction
M-01 · 2017

SVO

SVO 採半直接法（semi-direct）：以直接法追蹤並三角化影像梯度高的像素（含弱角點與邊緣），再以成熟的特徵式方法聯合最佳化結構與運動，並用顯式建模離群值的機率深度濾波器估計深度。期刊版將方法擴充到多相機、邊緣特徵、運動先驗及魚眼等大視角鏡頭。其定位為速度優先的視覺里程計，沒有迴圈閉合。

里程計＋局部建圖近十年已讀全文同儕審查已出版視覺、視覺慣性與 RGB-D 稠密 SLAM

閱讀版本
accepted manuscript (18 pages, UZH/ZORA copy TRO16_Forster-SVO.pdf) read in full; version of record IEEE T-RO 33(2):249-265 checked on IEEE Xplore (NTU access): Sec. XI to XIII text and Table I to III images read; VoR Table I differs from the manuscript (ORB-SLAM values replaced by values from the DSO paper and a real-time ORB-SLAM column added), so Table I values follow the VoR
全文取得途徑
國立臺灣大學機構訂閱
本頁內容

SVO tracks and triangulates high-gradient pixels directly but optimises structure and motion with feature-based methods, using a robust depth filter; the journal version adds multi-camera, edgelet, motion-prior and wide-FoV support.

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、multi-camera、fisheye
原文測試平台UAV、handheld
狀態估計three-step motion estimation: coarse-to-fine sparse image alignment on 4x4 patches with a robust cost for frame-to-frame motion, 2D alignment of 8x8 feature patches against the frame of first observation, then reprojection-error refinement of the latest pose and points, or full keyframe bundle adjustment with iSAM2 (Sec. V, X-B, X-C, XI-B)
資料關聯semi-direct: direct alignment of high-gradient pixels (FAST corners, or the highest-gradient pixel as an edgelet in cells of 32x32 pixels without corners) plus feature alignment and refinement; depth by epipolar ZMSSD search on 8x8 patches feeding a Gaussian plus uniform depth filter with Beta inlier ratio; at most 180 matched features per frame (Sec. VI, X-C, X-D)
時間表示discrete poses
去畸變不適用
迴圈閉合none (visual odometry)
全域最佳化none
地圖表示sparse 3D points and edgelets with depth filters; only a small local map of the last five to ten keyframes is kept (Sec. XI-B-1)
先驗資訊optional relative translation prior (e.g. constant velocity) and relative rotation prior (e.g. integrated gyroscope) added to the alignment cost; known camera intrinsics and extrinsics from prior calibration (Sec. IV, IX)
可輸出幾何camera trajectory and sparse points
計算需求laptop Intel Core i7-2760QM (2.80 GHz): SVO Mono 2.53 ms per frame and about 55 % CPU at 20 Hz input versus 29.81 ms for ORB-SLAM and 23.23 ms for LSD-SLAM (Table II); per-component timings also reported for an NVIDIA Jetson TX1 (Table III)

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
相機down-facing perspective camera (model not reported)方法輸入own circle datasetcircle flight on a micro aerial vehicle(Forster et al., 2017b, Sec. XI-B-4)
相機wide fisheye camera (model not reported)方法輸入own circle datasetsame circle trajectory flown again with a fisheye lens(Forster et al., 2017b, Sec. XI-B-4)
雙目相機VI Sensor歸入：VI-Sensor資料集感測器EuRoCstereo images and inertial data; mounted on a micro aerial vehicle; 11 sequences, 19 min in three indoor environments(Forster et al., 2017b, Sec. XI-B-1)
RGB-D 相機Microsoft Kinect RGB-D camera資料集感測器TUM RGB-Dimages of worse quality than the VI-Sensor (rolling shutter, motion blur); fr2_desk 18.8 m and fr2_xyz 7 m trajectories; SVO is not among the methods marked as using the depth sensor in Table IV(Forster et al., 2017b, Sec. XI-B-2, Table IV)
全測站Leica MS50 laser tracking system參考或真值量測EuRoCprovides the ground-truth trajectory(Forster et al., 2017b, Sec. XI-B-1)
載具平台micro aerial vehicle (model not reported)方法輸入own circle datasetflown in a motion capture room along a commanded circle(Forster et al., 2017b, Sec. XI-B-4)
運算硬體Intel Core i7-2760QM laptop執行運算平台未標示2.80 GHz(Forster et al., 2017b, Sec. XI-B-1, Table II)
運算硬體NVIDIA Jetson TX1執行運算平台未標示ARM processor; component timings of SVO Mono(Forster et al., 2017b, Table III)
其他motion capture system (model not reported)參考或真值量測TUM RGB-Dground truth for TUM RGB-D(Forster et al., 2017b, Sec. XI-B-2)

作者報告的優勢與限制

優勢

  • Significantly faster than state-of-the-art VO while competitive in accuracy (abstract)
  • Tracks weak corners and edgelets in low- or high-frequency texture (abstract)

限制

  • keep; add: monocular SVO fails on several EuRoC Vicon room sequences because of abrupt illumination changes and on-spot rotations (Sec. XI-B-1, Table I); iSAM2 bundle adjustment could not be applied on ICL-NUIM because frequent on-spot rotations leave underconstrained variables (Sec. XI-B-3)

營建工程相關證據

論文未報告營建測試；作者提及四旋翼視覺飛行與手機 3D 掃描應用，但未提供工地證據。

原文驗證環境：公開基準、模擬、受控實驗

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 11 個比較組，合計 154 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 7 組列在最後，並連到性能比較頁。

Forster et al., 2017b · Table I 本方法 88 筆

指標absolute translation error (RMSE)

表格設定（擷取紀錄原文）：EuRoC; absolute translation error RMSE of keyframe positions after least-squares translation and scale alignment, averaged over five runs; loop closure deactivated for ORB-SLAM and LSD-SLAM; ORB-SLAM and DSO values taken from the DSO paper [42] with and without enforced real-time execution; x = tracking failed (version of record Table I) (Forster et al., 2017b, Table I)

absolute translation error (RMSE)，EuRoC · Machine Hall 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。

這些是 Forster et al., 2017b 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。

統計量：均方根誤差（RMSE）；對齊方式：Sim(3) 相似對齊；單位：m；場景：indoor machine hall, micro aerial vehicle

資料來源作者報告值(Forster et al., 2017b, Table I)

數值與出處
方法（原文寫法）報告值出處
SVO (stereo)本方法原文提出0.08 m(Forster et al., 2017b, Table I)
SVO (stereo, edgelets)本方法原文提出0.08 m(Forster et al., 2017b, Table I)
SVO (stereo, edgelets + prior)本方法原文提出0.04 m(Forster et al., 2017b, Table I)
SVO (stereo, bundle adjustment)本方法原文提出0.04 m(Forster et al., 2017b, Table I)
SVO (monocular)本方法原文提出0.17 m(Forster et al., 2017b, Table I)
SVO (monocular, edgelets)本方法原文提出0.17 m(Forster et al., 2017b, Table I)
SVO (monocular, edgelets + prior)本方法原文提出0.1 m(Forster et al., 2017b, Table I)
SVO (monocular, bundle adjustment)本方法原文提出0.06 m(Forster et al., 2017b, Table I)
ORB-SLAM (monocular, no loop-closure)0.02 m(Forster et al., 2017b, Table I)
ORB-SLAM (monocular, no loop, real-time)0.61 m(Forster et al., 2017b, Table I)
DSO (monocular)0.05 m(Forster et al., 2017b, Table I)
DSO (monocular, real-time)0.05 m(Forster et al., 2017b, Table I)
LSD-SLAM (monocular, no loop-closure)0.18 m(Forster et al., 2017b, Table I)

Forster et al., 2017b · Table II 本方法 24 筆

資料集與序列EuRoC · Machine Hall 01

表格設定（擷取紀錄原文）：Processing time per frame and average CPU load at a constant 20 Hz input, averaged over three runs of EuRoC Machine Hall 01; all algorithms multi-threaded; CPU load reported with a +/- spread (e.g. 55 +/- 10 %); values identical in the accepted manuscript and the version-of-record table image (Forster et al., 2017b, Table II)

processing time per frame (mean)，EuRoC · Machine Hall 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。

這些是 Forster et al., 2017b 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。

統計量：平均值（mean）；對齊方式：未對齊；單位：ms；場景：indoor machine hall

資料來源作者報告值(Forster et al., 2017b, Table II)

數值與出處
方法（原文寫法）報告值出處
SVO Mono本方法原文提出硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)2.53 ms(Forster et al., 2017b, Table II)
SVO Mono + Prior本方法原文提出硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)2.32 ms(Forster et al., 2017b, Table II)
SVO Mono + Prior + Edgelet本方法原文提出硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)2.51 ms(Forster et al., 2017b, Table II)
SVO Mono + Bundle Adjustment本方法原文提出硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)5.25 ms(Forster et al., 2017b, Table II)
SVO Stereo本方法原文提出硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)4.7 ms(Forster et al., 2017b, Table II)
SVO Stereo + Prior本方法原文提出硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)3.86 ms(Forster et al., 2017b, Table II)
SVO Stereo + Prior + Edgelet本方法原文提出硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)4.12 ms(Forster et al., 2017b, Table II)
SVO Stereo + Bundle Adjustment本方法原文提出硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)7.61 ms(Forster et al., 2017b, Table II)
ORB Mono SLAM (No loop closure)硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)29.81 ms(Forster et al., 2017b, Table II)
LSD Mono SLAM (No loop closure)硬體：laptop with Intel Core i7-2760QM (2.80 GHz); images supplied at 20 Hz (Sec. XI-B-1)23.23 ms(Forster et al., 2017b, Table II)

Teed et al., 2023 · Table 2 本方法 12 筆

指標ATE[m]

表格設定（擷取紀錄原文）：EuRoC MAV monocular VO, ATE[m] after similarity alignment with EVO (Appendix C); median of 5 runs for DPVO; every other frame skipped (Teed et al., 2023, Table 2)

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。

這些是 Teed 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。

統計量：原文未報告；對齊方式：Sim(3) 相似對齊；單位：m；場景：indoor machine hall and Vicon room (MAV)

資料來源作者報告值(Teed et al., 2023, Table 2)

數值與出處
方法（原文寫法）報告值出處
TartanVO [ 43 ]0.639 m(Teed et al., 2023, Table 2)
SVO [ 15 ]本方法0.1 m(Teed et al., 2023, Table 2)
DSO [ 12 ]0.046 m(Teed et al., 2023, Table 2)
DROID-VO [ 37 ]0.163 m(Teed et al., 2023, Table 2)
Ours (Default)原文提出0.087 m(Teed et al., 2023, Table 2)
Ours (Fast)原文提出0.101 m(Teed et al., 2023, Table 2)

Deng & Gan, 2026 · Table 11 本方法 10 筆

指標ATE RMSE [m]

表格設定（擷取紀錄原文）：EuRoC MH01 to MH05; input modality for the 3DGS methods not stated; classical-method values match those listed in DROID-SLAM Tables 3 and 5 after rounding (Deng & Gan, 2026, Table 11)

ATE RMSE [m]，EuRoC MAV · MH01 Easy

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。

這些是 Deng & Gan, 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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：m；場景：large-scale indoor sequences with fast camera motion (Sec. 4.6)

資料來源作者報告值(Deng & Gan, 2026, Table 11)

數值與出處
方法（原文寫法）報告值出處
DSO (monocular, w/o loop)0.05 m(Deng & Gan, 2026, Table 11)
SVO (monocular, w/o loop)本方法0.1 m(Deng & Gan, 2026, Table 11)
ORB-SLAM (monocular, with loop)0.07 m(Deng & Gan, 2026, Table 11)
DSM (monocular, w/o loop)0.04 m(Deng & Gan, 2026, Table 11)
VINS-Fusion (stereo, with loop)0.54 m(Deng & Gan, 2026, Table 11)
SVO (stereo, w/o loop)本方法0.04 m(Deng & Gan, 2026, Table 11)
ORB-SLAM3 (stereo, with loop)0.03 m(Deng & Gan, 2026, Table 11)
SplaTAM (w/o loop)0.39 m(Deng & Gan, 2026, Table 11)
GI-SLAM (w/o loop)0.1 m(Deng & Gan, 2026, Table 11)
Baseline MonoGS (w/o loop)0.12 m(Deng & Gan, 2026, Table 11)
MCGS-SLAM (Ours, w/o loop)原文提出0.05 m(Deng & Gan, 2026, Table 11)

其他比較組

列出其餘 7 個比較組

來源

  • Forster et al., 2017b

    Christian Forster, Zichao Zhang, Michael Gassner, Manuel Werlberger, Davide Scaramuzza（2017）SVO: Semidirect Visual Odometry for Monocular and Multicamera SystemsIEEE Transactions on Robotics, 33(2):249-265

    DOI 10.1109/tro.2016.2623335程式碼

    同儕審查已出版已讀全文近十年查證後修正

  • 會議版：SVO: Fast semi-direct monocular visual odometry (ICRA 2014, pp. 15-22) 10.1109/ICRA.2014.6906584
  • 程式碼釋出：rpg_svo (GPLv3 per README; README cites the ICRA 2014 paper, i.e., the conference-version implementation) https://github.com/uzh-rpg/rpg_svo
  • 程式碼釋出：rpg_svo_pro_open (SVO Pro; README calls it the newest SVO version, cites the T-RO 2017 paper, and adds a sliding-window VIO back-end modified from OKVIS, an iSAM2 global map and DBoW2 loop closure, i.e., a superset of the paper's VO) https://github.com/uzh-rpg/rpg_svo_pro_open

程式碼：https://github.com/uzh-rpg/rpg_svo_pro_open（授權：GPLv3 (LICENSE file of rpg_svo_pro_open)）。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。

回到方法圖鑑

SLAM 點雲圖鑑

本站為綜述論文《SLAM-Based Point Clouds for Construction: A Review》研究過程的互動說明文件。互動圖為教學模擬，不代表任何特定系統的實測表現；方法描述、數值與結論以所引原始文獻為準。

內容基準日 2026-09-25 · 研究與網站製作使用 AI 輔助工具，見 證據標記與方法

選擇開啟Esc關閉