LSD-SLAM
LSD-SLAM 為直接法(direct method)單眼 SLAM,不萃取特徵點,而是對影像梯度明顯的像素做光度誤差對齊,並以許多小基線立體比對濾波估計關鍵影格的半稠密(semi-dense)深度圖。新關鍵影格以 sim(3) 直接對齊連接鄰近關鍵影格,以明確偵測尺度漂移,並在位姿圖上做全域最佳化。地圖可輸出為半稠密點雲,但尺度仍非公制。
本頁內容
LSD-SLAM tracks a monocular camera by direct photometric alignment and builds a Sim(3) pose graph of keyframes carrying filtered semi-dense depth maps, explicitly modelling scale drift.
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 | handheld |
| 狀態估計 | weighted Gauss-Newton direct image alignment on se(3)/sim(3); keyframe pose-graph optimisation with Sim(3) edges (Sec. 2.2, 3.1) |
| 資料關聯 | direct photometric alignment on high-gradient (semi-dense) pixels; depth by filtering many small-baseline stereo comparisons (abstract; Sec. 3.1) |
| 時間表示 | discrete poses (keyframes) |
| 去畸變 | 不適用 |
| 迴圈閉合 | candidates from the ten closest keyframes plus an appearance-based candidate (OpenFABMAP, ref. [11]), accepted after a reciprocal sim(3) tracking consistency check (Sec. 3.5 Constraint Acquisition) |
| 全域最佳化 | Sim(3) pose-graph optimisation of keyframes, run continuously in the background (Sec. 3.6 Map optimization; Fig. 3; conclusion) |
| 地圖表示 | Pose graph of keyframes; each keyframe stores the image, a semi-dense inverse depth map and its variance defined only near sufficiently large intensity gradients, scaled to mean inverse depth one; edges hold sim(3) transforms with covariance; the map can be exported as point clouds |
| 先驗資訊 | No prior map. The method bootstraps from a first keyframe with random depth and large variance and needs sufficient translational motion in the first seconds; for the TUM RGB-D evaluation the very first sensor depth map was used to bootstrap and obtain the correct initial scale |
| 可輸出幾何 | semi-dense point cloud from keyframe depth maps and poses (conclusion); monocular, scale not metric |
| 計算需求 | Real time on a CPU; the pipeline figure assumes 640x480 images at 30 Hz and map optimisation runs continuously in the background; no CPU model or timing numbers are reported (an odometry-only variant is cited as running on a smartphone, ref. [22]) |
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 | hand-held monocular camera (model not stated) | 方法輸入 | 未標示 | pipeline figure assumes 640x480 at 30 Hz; used for the qualitative outdoor trajectories of about 500 m | (Engel et al., 2014, Sec. 4, 4.1; Fig. 3) |
| RGB-D 相機 | TUM RGB-D benchmark sensor (model not stated) | 資料集感測器 | TUM RGB-D | images used as monocular input; the first depth map used only for bootstrapping and initial scale | (Engel et al., 2014, Sec. 4.2) |
作者報告的優勢與限制
優勢
- Hand-held trajectories over 500 m with large scene-scale variation (conclusion)
- Real-time on CPU (abstract)
限制
- Direct sim(3) image alignment is non-convex and needs accurate initialisation, which matters for loop-closure constraints; the authors report that ESM and a very coarse pyramid start enlarge the convergence radius so that it sufficed in practice even for large-scale loop closures (Sec. 3.5; Sec. 4.3)
- ORB-SLAM authors note LSD-SLAM reduces map optimisation to a pose graph and discards sensor measurements (Mur-Artal et al., 2015, Sec. IX-B)
- Convergence of the random-depth initialisation is not thoroughly evaluated and is left to future work (Sec. 3.1)
營建工程相關證據
論文未報告營建測試;以手持序列及 TUM RGB-D 評估。半稠密點雲僅來自影像梯度區域且尺度未定,用於工程幾何時需外部尺度與參考控制(推論)。
原文驗證環境:公開基準、受控實驗、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 8 個比較組,合計 87 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 4 組列在最後,並連到性能比較頁。
Tateno et al., 2017 · Table 1 本方法 40 筆
表格設定(擷取紀錄原文):Percentage of key-frame depth values within 10% of ground-truth depth (accuracy and density together) (Tateno et al., 2017, Table 1)
Perc. Correct Depth (error < 10%),ICL-NUIM · office0
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 Tateno et al., 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(Tateno et al., 2017, Table 1)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| CNN-SLAM (Our Method)原文提出 | 19.41% | (Tateno et al., 2017, Table 1) |
| LSD-BS [4] (LSD-SLAM bootstrapped with ground-truth depth)本方法 | 0.603% | (Tateno et al., 2017, Table 1) |
| LSD [4] (LSD-SLAM)本方法 | 0.335% | (Tateno et al., 2017, Table 1) |
| ORB [20] (ORB-SLAM) | 0.018% | (Tateno et al., 2017, Table 1) |
| Laina [16] (CNN depth fed to point-based fusion) | 17.194% | (Tateno et al., 2017, Table 1) |
| Remode [23] (REMODE, poses from LSD-BS) | 4.479% | (Tateno et al., 2017, Table 1) |
Mur-Artal et al., 2015 · Table III 本方法 16 筆
指標Absolute KeyFrame Trajectory RMSE
表格設定(擷取紀錄原文):TUM RGB-D keyframe ATE RMSE (cm), median over 5 executions; ORB-SLAM, PTAM (two manually chosen initial frames) and LSD-SLAM (first 10 keyframes discarded) aligned with Sim(3); RGBD-SLAM trajectories from the benchmark website aligned with SE(3), Sim(3) value in brackets; X = tracking lost (Mur-Artal et al., 2015, Table III)
Absolute KeyFrame Trajectory RMSE,TUM RGB-D · fr1_xyz
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 et al., 2015 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 et al., 2015, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM原文提出 | 0.9 cm | (Mur-Artal et al., 2015, Table III) |
| PTAM | 1.15 cm | (Mur-Artal et al., 2015, Table III) |
| LSD-SLAM本方法 | 9 cm | (Mur-Artal et al., 2015, Table III) |
Forster et al., 2017b · Table I 本方法 11 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。
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 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。
資料來源作者報告值(Forster et al., 2017b, Table I)
| 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) | 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)本方法 | 0.18 m | (Forster et al., 2017b, Table I) |
Czarnowski et al., 2020 · Table II 本方法 8 筆
指標Perc. Correct Depth [%]
表格設定(擷取紀錄原文):Percentage of key-frame pixels with depth within 10% of ground truth; DeepFactors trajectory and depth scaled by the optimal scale from the TUM scripts; other values taken from the CNN-SLAM paper (Czarnowski et al., 2020, Table II)
Perc. Correct Depth [%],ICL-NUIM · office0
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 Czarnowski et al., 2020 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(Czarnowski et al., 2020, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| CNN-SLAM [ 18 ] | 19.41% | (Czarnowski et al., 2020, Table II) |
| LSD-BS [ 19 ]本方法 | 0.6% | (Czarnowski et al., 2020, Table II) |
| Laina [ 37 ] | 17.19% | (Czarnowski et al., 2020, Table II) |
其他比較組
來源
Engel et al., 2014
(2014)LSD-SLAM: Large-Scale Direct Monocular SLAMComputer Vision – ECCV 2014 (Lecture Notes in Computer Science), pp. 834-849
DOI 10.1007/978-3-319-10605-2_54程式碼
同儕審查已出版已讀全文經典查證後修正
相關版本
- 程式碼釋出:lsd_slam https://github.com/tum-vision/lsd_slam
程式碼:https://github.com/tum-vision/lsd_slam(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。