SLAM 點雲圖鑑Point clouds for construction
M-01 · 2018

DSO

DSO 是直接稀疏法的單眼視覺里程計（visual odometry），直接最小化光度誤差，並在滑動視窗內聯合最佳化相機位姿、相機內參、仿射亮度參數與逆深度，舊狀態以邊際化（marginalization）移除。它不使用平滑先驗，而是在影像中均勻取樣具梯度的像素，包括白牆上的弱梯度與邊緣，並整合曝光、暗角與非線性響應的光度校正。DSO 不含迴圈閉合，屬里程計而非完整 SLAM。

里程計＋局部建圖經典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
arXiv 1607.02565v2 (2016-10-07; v2 comment: corrected a bug that made ORB-SLAM real-time results worse, added refs [12], [13], [19] and Fig. 11, extended conclusion); IEEE TPAMI 40(3) version of record not compared
全文取得途徑
arXiv 預印本
本頁內容

DSO is a direct sparse monocular odometry that jointly optimises poses, intrinsics, affine brightness and inverse depths in a marginalised sliding window, sampling gradient pixels evenly and using full photometric calibration; it has no loop closure.

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
原文測試平台UAV (EuRoC MAV quadrocopter sequences, left and right images used separately)、TUM monoVO sequences, 50 photometrically calibrated indoor and outdoor videos (carrying mode not stated in this paper)、synthetic ray-traced ICL-NUIM sequences、Fig. 1 video recorded while cycling around a building
狀態估計sliding-window Gauss-Newton (up to 6 iterations per new keyframe, no Levenberg-Marquardt damping) jointly over poses, affine brightness parameters, inverse depths and camera intrinsics, with First Estimate Jacobians and Schur-complement marginalisation; window Nf = 7 keyframes and Np = 2000 active points; keyframes marginalised by a distance score and residuals that would break Hessian sparsity are dropped (about half of all residuals)
資料關聯direct photometric error of an 8-pixel residual pattern with Huber norm and gradient-dependent weighting, on points sampled with a region-adaptive gradient threshold (32x32 blocks, three passes with lower thresholds); inverse depth in a host frame; candidates tracked by discrete epipolar search before activation; new frames tracked by two-frame direct alignment to the newest keyframe's projected semi-dense depth map with a constant motion model, with up to 27 small-rotation retries on failure
時間表示discrete poses (keyframes)
去畸變不適用 (rolling shutter not modelled; simulated as low-frequency geometric noise in Sec. 4.3, where DSO degrades much faster than ORB-SLAM and, according to the authors, optimisation likely fails entirely for noise amplitude above 1.5 (alleviable with a coarser pyramid level); tight rolling-shutter modelling cited as remedy)
迴圈閉合none (visual odometry; explicit loop closure disabled for ORB-SLAM in comparisons for fairness, Sec. 4)
全域最佳化none
地圖表示sparse set of points with inverse depth in active keyframes
先驗資訊photometric calibration (exposure time, vignetting, response function) (abstract)
可輸出幾何point clouds accumulated from the odometry without loop closure, density set by the number of active points (Np = 500 to 10000 shown); monocular scale unobservable (scale is a null space of the energy) and evaluated with Sim(3) alignment and scale drift
計算需求CPU only; real time on a laptop; hard-enforced real-time evaluations on an Intel i7-4910MQ CPU; non-real-time evaluation in a sequentialised single-threaded mode about four times slower than real time on 20 dedicated workstations; reduced settings run at 5 times real time

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TUM monoVO camera (model not reported in this paper)資料集感測器TUM monoVO50 photometrically calibrated sequences (response, vignetting, exposure times), 105 minutes, about 190,000 frames; exposure varied from 0.018 to 10.5 ms in an indoor-outdoor sequence(Engel et al., 2018, Sec. 2.1.2, Fig. 3; Sec. 4 datasets; Fig. 11)
雙目相機EuRoC MAV stereo camera (model not reported in this paper)資料集感測器EuRoC MAV11 stereo-inertial sequences, 19 minutes; left and right videos used separately as monocular input; no photometric calibration or exposure times; shaky initialisation segments cropped(Engel et al., 2018, Sec. 4 datasets and Methodology)
運算硬體Intel i7-4910MQ CPU執行運算平台未標示hard-enforced real-time evaluation(Engel et al., 2018, Sec. 4.1)
運算硬體20 dedicated workstations (model not reported)執行運算平台未標示non-real-time sequentialised evaluation runs(Engel et al., 2018, Sec. 4 Methodology)

作者報告的優勢與限制

優勢

  • Can use pixels on edges and smooth intensity variations on mostly white walls (abstract)
  • Photometric calibration improves performance versus brightness constancy (Sec. 4.2, Fig. 15; Sec. 5)
  • With high settings produces semi-dense models similar in density to LSD-SLAM (Sec. 1.2)
  • Restricting candidates to FAST corners clearly reduces accuracy and robustness (Sec. 4.2, Fig. 17)
  • Slightly more robust than ORB-SLAM to strong photometric noise (simulated anisotropic blur) (Sec. 4.3, Fig. 21)
  • Outperforms ORB-SLAM in accuracy and robustness on TUM monoVO and ICL-NUIM (Sec. 4.1)

限制

  • Indirect (geometric-error) approaches are more robust to geometric noise such as poor intrinsic calibration or rolling shutter (Sec. 4.3; Sec. 5)
  • Using more points makes models denser but does not increase tracking accuracy (Sec. 4.2; Sec. 5)
  • No loop closure or relocalization; points and frames leaving the view are permanently marginalised (Sec. 4.1)
  • On EuRoC MAV ORB-SLAM is more accurate (but less robust), attributed to missing photometric calibration and many small loops (Sec. 4.1)
  • Fewer than about 4 keyframes per second reduces robustness and more than 15 reduces accuracy (Sec. 4.2)
  • Higher non-convexity of the photometric model likely restricts it to video processing (Sec. 5)

營建工程相關證據

論文未報告營建測試；評估資料為 TUM monoVO、EuRoC 與 ICL-NUIM。作者指出可利用白牆弱梯度，但對捲簾快門與內參誤差敏感，現場使用消費型相機時須注意（推論）。

原文驗證環境：公開基準、模擬

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 17 個比較組，合計 104 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 13 組列在最後，並連到性能比較頁。

Forster et al., 2017b · Table I 本方法 22 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)原文提出0.08 m(Forster et al., 2017b, Table I)
SVO (stereo, edgelets)原文提出0.08 m(Forster et al., 2017b, Table I)
SVO (stereo, edgelets + prior)原文提出0.04 m(Forster et al., 2017b, Table I)
SVO (stereo, bundle adjustment)原文提出0.04 m(Forster et al., 2017b, Table I)
SVO (monocular)原文提出0.17 m(Forster et al., 2017b, Table I)
SVO (monocular, edgelets)原文提出0.17 m(Forster et al., 2017b, Table I)
SVO (monocular, edgelets + prior)原文提出0.1 m(Forster et al., 2017b, Table I)
SVO (monocular, bundle adjustment)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)本方法0.05 m(Forster et al., 2017b, Table I)
DSO (monocular, real-time)本方法0.05 m(Forster et al., 2017b, Table I)
LSD-SLAM (monocular, no loop-closure)0.18 m(Forster et al., 2017b, Table I)

Teed et al., 2023 · Table 3 本方法 20 筆

指標ATE

表格設定（擷取紀錄原文）：TUM RGB-D freiburg1 monocular VO, ATE; x = method failed and output no trajectory; '-' = average not computed; median of 5 trials (Teed et al., 2023, Table 3)

ATE，TUM RGB-D · fr1/360

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。

  • 失敗

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：原文未報告；場景：indoor, erratic motion and motion blur

資料來源作者報告值(Teed et al., 2023, Table 3)

數值與出處
方法（原文寫法）報告值出處
ORB-SLAM3 [ 27 ]無數值失敗註記（擷取紀錄）：failed(Teed et al., 2023, Table 3)
DSO [ 12 ]本方法0.173(Teed et al., 2023, Table 3)
DSO-Realtime [ 12 ]本方法0.172(Teed et al., 2023, Table 3)
DROID-VO [ 37 ]0.161(Teed et al., 2023, Table 3)
Ours (Default)原文提出0.135(Teed et al., 2023, Table 3)
Ours (Fast)原文提出0.169(Teed et al., 2023, Table 3)

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 ]0.1 m(Teed et al., 2023, Table 2)
DSO [ 12 ]本方法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)

Gao et al., 2018 · Table I 本方法 11 筆

指標ATE error (m)

表格設定（擷取紀錄原文）：KITTI Odometry training sequences, monocular setting; ATE (m) after Sim(3) alignment to ground truth; x = failure (Gao et al., 2018, Table I)

ATE error (m)，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。

這些是 Gao et al., 2018 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：KITTI Odometry training sequences (platform and scene not described in the paper)

資料來源作者報告值(Gao et al., 2018, Table I)

數值與出處
方法（原文寫法）報告值出處
Mono DSO本方法126.7 m(Gao et al., 2018, Table I)
LDSO原文提出9.322 m(Gao et al., 2018, Table I)
ORB-SLAM2 (monocular)8.27 m(Gao et al., 2018, Table I)

其他比較組

列出其餘 13 個比較組

來源

程式碼：https://github.com/JakobEngel/dso（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。

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