ManhattanSLAM
ManhattanSLAM 是只用 CPU 的室內 RGB-D SLAM。每一影格擷取 ORB 點、LSD 線段與深度圖中的平面；只要找到兩或三個相互垂直的平面就組成一個曼哈頓座標系（Manhattan Frame），並把場景視為多個曼哈頓座標系的混合。若目前觀測到的曼哈頓座標系先前已存入曼哈頓地圖，就直接由兩次觀測求出無漂移的旋轉，平移再由點、線、面特徵最佳化；若場景不符合曼哈頓假設，則以點、線、面及平行與垂直平面約束估計完整六自由度位姿。稠密建圖沿用 Dense Surfel Mapping 的超像素面元，但平面區域改以稀疏地圖中的平面點建立面元，以節省記憶體。
本頁內容
CPU RGB-D SLAM for structured interiors: points, lines and planes are tracked; Manhattan frames detected from mutually perpendicular planes give drift-free rotation when revisited (translation then from features), with full point-line-plane optimization in non-Manhattan scenes; dense surfels reuse sparse-map planes for planar regions and superpixel surfels elsewhere.
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 | RGB-D camera (synthetic ICL-NUIM, TUM RGB-D and TAMU RGB-D sequences; sensor models not named in the paper) |
|---|---|
| 原文測試平台 | real RGB-D sequences (TUM RGB-D; TAMU RGB-D long indoor loop sequences; capture platform not described in the paper)、simulation (ICL-NUIM living room and office) |
| 狀態估計 | Keyframe-based feature SLAM with constant-velocity prediction and local-map refinement; in Manhattan scenes rotation is taken drift-free from a previously stored Manhattan-frame observation and only translation is optimized from point, line and plane errors; otherwise full 6-DoF Levenberg-Marquardt with Huber cost over point, line and plane reprojection errors plus parallel and perpendicular plane constraints |
| 資料關聯 | ORB points matched by projection and Hamming distance; LSD line segments matched with LBD descriptors; planes extracted by AHC from the downsampled point cloud and matched by normal angle and point-plane distance; Manhattan frames detected from two or three mutually perpendicular plane normals (SVD-orthogonalized) and matched through shared map-plane IDs |
| 時間表示 | discrete poses |
| 去畸變 | 不適用 (RGB-D input) |
| 迴圈閉合 | none (adding a loop-closure module is listed as future work) |
| 全域最佳化 | none reported; drift is limited by Manhattan-frame rotation estimates |
| 地圖表示 | sparse map of point, line and plane landmarks with keyframe co-visibility graph and a Manhattan map of MF observations; dense surfel map in which planar regions reuse sparse-map plane points (surfel radius from the 0.2 m downsampling voxel) and non-planar regions use superpixel surfels as in Dense Surfel Mapping |
| 先驗資訊 | Manhattan or mixture-of-Manhattan-frames structural assumption where detected; not required |
| 可輸出幾何 | trajectory, sparse point-line-plane map, dense surfel point cloud |
| 計算需求 | CPU only; Intel Core i5-8250U at 1.60 GHz x 8 with 19.5 GB RAM; about 15 Hz, 67 ms tracking and 40 ms superpixel extraction plus surfel fusion on a separate thread (Sec. IV) |
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 |
|---|---|---|---|---|---|
| 運算硬體 | Intel Core i5-8250U CPU @ 1.60GHz x 8, 19.5 GB RAM | 執行運算平台 | 未標示 | no GPU used; about 15 Hz | (Yunus et al., 2021, Sec. IV) |
作者報告的優勢與限制
優勢
- Lowest average ATE RMSE on the eight ICL-NUIM sequences (0.014 m) among the compared feature-based and Manhattan-based systems (Table I)
- Keeps tracking on cluttered TUM fr1 and fr2 sequences where the Manhattan-only methods fail (Table I)
- Loop-end drift on TAMU RGB-D of 0.53 m and 0.39 m versus 3.13 m and 2.22 m for ORB-SLAM2 without bundle adjustment and loop closure; Manhattan-frame tracking lowers drift from 0.77 to 0.53 m and from 0.81 to 0.39 m compared with feature tracking only (Table II)
- Dense reconstruction error 0.5 to 0.7 cm on ICL-NUIM living room, best on three of four sequences, on a CPU (Table III)
限制
- No loop-closure module (Sec. V)
- Noisy real depth degrades plane normals, so Manhattan-frame tracking is used less and the system falls back to feature tracking (Sec. IV-A2, IV-B)
- Planar surfel radius is fixed from the voxel size and does not follow the plane boundary (Sec. V)
- TAMU drift is measured only as the start-to-end distance of a loop because no ground truth exists (Sec. IV-B)
- ORB-SLAM2 and SP-SLAM were run with bundle adjustment and loop closure disabled, so the comparison excludes their global correction (Sec. IV)
營建工程相關證據
論文未在施工現場測試，評估限於 ICL-NUIM 合成室內、TUM RGB-D，以及 TAMU RGB-D 的室內走廊與大廳長序列。以相互垂直平面求無漂移旋轉的作法適合牆、樓板與天花板占多數的建築室內；但未完工或雜亂的施工現場可能不符合曼哈頓假設，此時系統會退回一般特徵追蹤（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 4 個比較組，合計 29 筆紀錄。
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)本方法原文提出 | 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) | 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) |
Yunus et al., 2021 · Table II 本方法 4 筆
指標accumulated drift (m)
表格設定（擷取紀錄原文）：TAMU RGB-D long indoor loops without ground truth; drift = Euclidean distance between start and end of the estimated loop trajectory (Yunus et al., 2021, Table II)
accumulated drift (m)，TAMU RGB-D · Corridor-A
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 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 II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Ours (ManhattanSLAM)本方法原文提出 | 0.53 m | (Yunus et al., 2021, Table II) |
| Ours/-MF (ablation: feature tracking only)本方法 | 0.77 m | (Yunus et al., 2021, Table II) |
| ORB-SLAM2 [6] (BA and loop closure disabled) | 3.13 m | (Yunus et al., 2021, Table II) |
Yunus et al., 2021 · Table III 本方法 4 筆
指標reconstruction error (cm)
表格設定（擷取紀錄原文）：ICL-NUIM living room; reconstruction error of the point cloud generated from the surfels (cm); ElasticFusion and InfiniTAM need a GPU, DSM and ManhattanSLAM run on CPU; comparator provenance not stated (Yunus et al., 2021, Table III)
reconstruction error (cm)，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。
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 III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| E-Fus [15] (ElasticFusion, IJRR version) | 0.7 cm | (Yunus et al., 2021, Table III) |
| InfiniTAM [41] (InfiniTAM v3 report) | 1.3 cm | (Yunus et al., 2021, Table III) |
| DSM [14] (Dense Surfel Mapping) | 0.7 cm | (Yunus et al., 2021, Table III) |
| Ours (ManhattanSLAM)本方法原文提出 | 0.5 cm | (Yunus et al., 2021, Table III) |
Yunus et al., 2021 · Text Sec.IV 本方法 3 筆
資料集與序列ICL-NUIM, TUM RGB-D and TAMU RGB-D
表格設定（擷取紀錄原文）：Average timing over the experiments; dense mapping runs on a separate thread (Yunus et al., 2021, Text Sec.IV)
runs at around 15 Hz，ICL-NUIM, TUM RGB-D and TAMU RGB-D
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 Yunus et al., 2021 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 |
|---|---|---|
| Ours (ManhattanSLAM)本方法原文提出硬體：Intel Core i5-8250U @ 1.60 GHz x 8, 19.5 GB RAM, no GPU | 15 Hz有附註註記（擷取紀錄）：other: approximate value ('around 15 Hz', Sec. IV) | (Yunus et al., 2021, Sec. IV) |
來源
Yunus et al., 2021
（2021）ManhattanSLAM: Robust Planar Tracking and Mapping Leveraging Mixture of Manhattan Frames2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 6687-6693
DOI 10.1109/icra48506.2021.9562030arXiv 2103.15068程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本：ManhattanSLAM (arXiv v1) https://arxiv.org/abs/2103.15068
- 程式碼釋出：razayunus/ManhattanSLAM (released after the arXiv v1, which states the code will be open-sourced in the future) https://github.com/razayunus/ManhattanSLAM
程式碼：https://github.com/razayunus/ManhattanSLAM（授權：GPL-3.0 (LICENSE.txt checked; closed-source commercial licence on request)）。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。