MASt3R-SLAM
MASt3R-SLAM 以 MASt3R 雙視角重建先驗為核心建構即時單目稠密 SLAM，只假設單一相機中心，用迭代投影做點圖匹配、以 Sim(3) 位姿處理預測間不一致的尺度，並以影像檢索做迴圈閉合與重定位，後端為二階全域最佳化。幾何評估在 EuRoC Vicon 房間以軌跡對齊結構掃描真值，在 7-Scenes 以 ICP 對齊深度反投影參考，並報告 RMSE。作者承認全域最佳化不精修幾何，且畸變大的相機會降低預測品質。
本頁內容
Real-time monocular dense SLAM built on MASt3R priors with Sim(3) poses, ray-error optimization, retrieval-based loop closure and dense pointmap fusion.
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 (uncalibrated, generic central camera) |
|---|---|
| 原文測試平台 | 未記錄 |
| 狀態估計 | Gauss-Newton second-order optimization of Sim(3) keyframe poses minimizing ray error; sparse Cholesky backend in CUDA |
| 資料關聯 | MASt3R pointmap matching via iterative projective ray search |
| 時間表示 | discrete poses |
| 去畸變 | 不適用 |
| 迴圈閉合 | incremental ASMK image retrieval + MASt3R decoder matching; relocalisation via retrieval |
| 全域最佳化 | second-order global optimization over the keyframe graph (first 7-DoF pose fixed) |
| 地圖表示 | per-keyframe canonical pointmaps with local fusion |
| 先驗資訊 | MASt3R learned two-view 3D reconstruction prior |
| 可輸出幾何 | dense point cloud from fused pointmaps + trajectory (scale via Sim(3), not guaranteed metric) |
| 計算需求 | Intel Core i9-12900K 3.50 GHz + NVIDIA GeForce RTX 4090; single-threaded system at about 15 FPS, so datasets were subsampled every 2 frames to simulate real time (not for ETH3D); average per-frame tracking 45.9 ms, per-keyframe backend 164.9 ms, 14.6 FPS; MASt3R encoder and decoder take about 64% of runtime; MASt3R outputs resized to 512 px on the largest side (Sec. 4, Supp. Sec. 10, Table 8) |
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 | monocular RGB camera (model not named) | 方法輸入 | TUM RGB-D; 7-Scenes; ETH3D-SLAM; EuRoC | monocular RGB input with no parametric camera model assumed beyond a unique camera centre; EuRoC images undistorted for the uncalibrated run | (Murai et al., 2025, Abstract; Sec. 3.1; Sec. 4; Sec. 4.1) |
| RGB-D 相機 | depth camera (model not named) | 參考或真值量測 | 7-Scenes | depth images back-projected with dataset poses to form the reference cloud; default factory intrinsics | (Murai et al., 2025, Sec. 4.2) |
| 運算硬體 | Intel Core i9 12900K 3.50GHz | 執行運算平台 | 未標示 | desktop CPU | (Murai et al., 2025, Sec. 4) |
| 運算硬體 | NVIDIA GeForce RTX 4090 | 執行運算平台 | 未標示 | single GPU | (Murai et al., 2025, Sec. 4) |
| 其他 | Vicon | 參考或真值量測 | EuRoC | Vicon trajectory; the estimated trajectory is aligned to it to place the estimated cloud in the frame of the EuRoC 3D structure-scan ground truth | (Murai et al., 2025, Sec. 4.2) |
作者報告的優勢與限制
優勢
- Lower Chamfer and accuracy RMSE than DROID-SLAM on EuRoC Vicon rooms (Chamfer 0.085 vs 0.117 m) and 7-Scenes seq-01 (0.066 vs 0.077 m); DROID-SLAM has lower ATE on EuRoC (0.022 vs 0.041 m) and better completion, whereas on 7-Scenes the calibrated ATE is similar (0.047 vs 0.049 m) (Sec. 4.2; Table 3)
- Works without known calibration (abstract)
- TUM RGB-D calibrated average ATE 0.030 m, lowest in Table 1; uncalibrated 0.060 m vs 0.158 m for DROID-SLAM with GeoCalib intrinsics (Table 1)
- ETH3D-SLAM train: lowest mean ATE (0.086 m) and highest AUC (23.935) among the monocular systems run (Fig. 5)
- Projective pointmap matching takes about 2 ms versus 2000 ms for MASt3R matching over all pixels (Table 4)
限制
- Geometry not refined in the global optimization (Sec. 5)
- Predictions degrade with lens distortion; trained on pinhole images (Sec. 5)
- Full-resolution decoder is a throughput bottleneck (Sec. 5)
- EuRoC average ATE 0.041 m is worse than DROID-SLAM and DPV-SLAM-based systems (0.022 to 0.024 m); authors note DROID-SLAM trains with greyscale augmentation (Sec. 4.1, Table 9)
- Uncalibrated EuRoC runs required undistorted images because MASt3R had not been trained on such distortion (Sec. 4.1)
- ETH3D evaluated only on train sequences because the official thresholds are too strict for monocular methods (Sec. 4.1)
- Without loop closure EuRoC Vicon ATE rises from 0.029 to 0.233 m, showing drift from biased MASt3R outputs (Table 7)
營建工程相關證據
論文未涉及營建場域；資料為 TUM RGB-D、7-Scenes、ETH3D-SLAM、EuRoC。
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 9 個比較組，合計 78 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 5 組列在最後，並連到性能比較頁。
Murai et al., 2025 · Table 2 本方法 16 筆
指標ATE (m)
表格設定（擷取紀錄原文）：ATE RMSE (m) on 7-Scenes, monocular RGB, scaled trajectory alignment; sequences follow NICER-SLAM; NICER-SLAM values reported from NICER-SLAM; frames subsampled every 2 to simulate real time. (Murai et al., 2025, Table 2)
ATE (m)，7-Scenes · chess
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 Murai 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(Murai et al., 2025, Table 2)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| NICER-SLAM | 0.033 m | (Murai et al., 2025, Table 2) |
| DROID-SLAM | 0.036 m | (Murai et al., 2025, Table 2) |
| MASt3R-SLAM (Ours, calibrated)本方法原文提出 | 0.053 m | (Murai et al., 2025, Table 2) |
| MASt3R-SLAM (Ours*, uncalibrated)本方法原文提出 | 0.063 m | (Murai et al., 2025, Table 2) |
Murai et al., 2025 · Table 3 本方法 16 筆
表格設定（擷取紀錄原文）：Reconstruction evaluation (m). Accuracy and completion are RMSE of nearest-neighbour distances with a 0.5 m maximum distance, Chamfer is their average; no estimated points filtered; unobservable reference points removed. 7-Scenes uses seq-01 of each scene with a reference cloud back-projected from depth images and aligned to the estimate by ICP (scale handling not stated); EuRoC geometry uses the Vicon room sequences with the estimate aligned via its trajectory to Vicon (scaled alignment). Ours* = without known calibration; Spann3R keyframe every 20 or 2 images; Spann3R excluded on EuRoC. (reviewer inference) The EuRoC ATE column equals the 11-sequence averages of Table 9 (0.022, 0.041, 0.164), not a Vicon-only average. (Murai et al., 2025, Table 3)
ATE，7-Scenes · seq-01 of each scene
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。
- 不適用
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 Murai 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(Murai et al., 2025, Table 3)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| DROID-SLAM | 0.049 m | (Murai et al., 2025, Table 3) |
| Spann3R @20 | 無數值不適用註記（擷取紀錄）：不適用 (N/A) | (Murai et al., 2025, Table 3) |
| Spann3R @2 | 無數值不適用註記（擷取紀錄）：不適用 (N/A) | (Murai et al., 2025, Table 3) |
| MASt3R-SLAM (Ours, calibrated)本方法原文提出 | 0.047 m | (Murai et al., 2025, Table 3) |
| MASt3R-SLAM (Ours*, uncalibrated)本方法原文提出 | 0.066 m | (Murai et al., 2025, Table 3) |
Maggio et al., 2025 · Table 1 本方法 16 筆
指標ATE RMSE [m]
表格設定（擷取紀錄原文，這些數值分屬表中不同部分）：(Maggio et al., 2025, Table 1)
- ATE RMSE on 7-Scenes computed with evo (alignment not stated); calibrated intrinsics; value reported from MASt3R-SLAM
- ATE RMSE on 7-Scenes computed with evo (alignment not stated); uncalibrated; DROID-SLAM* intrinsics from an automatic calibration pipeline, run by the authors
- ATE RMSE on 7-Scenes computed with evo (alignment not stated); uncalibrated; value reported from MASt3R-SLAM
- ATE RMSE on 7-Scenes computed with evo (alignment not stated); uncalibrated; VGGT-SLAM average of five runs
ATE RMSE [m]，7-Scenes · chess
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 Maggio 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(Maggio et al., 2025, Table 1)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| NICER-SLAM | 0.033 m | (Maggio et al., 2025, Table 1) |
| DROID-SLAM | 0.036 m | (Maggio et al., 2025, Table 1) |
| MASt3R-SLAM本方法 | 0.053 m | (Maggio et al., 2025, Table 1) |
| DROID-SLAM* | 0.047 m | (Maggio et al., 2025, Table 1) |
| MASt3R-SLAM*本方法 | 0.063 m | (Maggio et al., 2025, Table 1) |
| Ours (Sim(3), w = 32) | 0.037 m | (Maggio et al., 2025, Table 1) |
| Ours (SL(4), w = 32)原文提出 | 0.036 m | (Maggio et al., 2025, Table 1) |
Maggio et al., 2025 · Table 2 本方法 11 筆
資料集與序列TUM RGB-D · Avg
表格設定（擷取紀錄原文）：ATE RMSE on TUM RGB-D; calibrated baselines with values reported from MASt3R-SLAM; only the average column extracted (row cap) (Maggio et al., 2025, Table 2)
ATE RMSE [m], average over 9 sequences，TUM RGB-D · Avg
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。
- 不適用
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 Maggio 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(Maggio et al., 2025, Table 2)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM3 | 無數值不適用註記（擷取紀錄）：不適用 (average reported as N/A; cells for 360, floor, room, rpy and teddy marked x without explanation in the paper) | (Maggio et al., 2025, Table 2) |
| DeepV2D | 0.375 m | (Maggio et al., 2025, Table 2) |
| DeepFactors | 0.233 m | (Maggio et al., 2025, Table 2) |
| DPV-SLAM | 0.076 m | (Maggio et al., 2025, Table 2) |
| DPV-SLAM++ | 0.054 m | (Maggio et al., 2025, Table 2) |
| GO-SLAM | 0.035 m | (Maggio et al., 2025, Table 2) |
| DROID-SLAM | 0.038 m | (Maggio et al., 2025, Table 2) |
| MASt3R-SLAM本方法 | 0.03 m | (Maggio et al., 2025, Table 2) |
其他比較組
來源
Murai et al., 2025
（2025）MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 16695-16705
DOI 10.1109/cvpr52734.2025.01556arXiv 2412.12392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：arXiv:2412.12392 https://arxiv.org/abs/2412.12392
程式碼：https://github.com/rmurai0610/MASt3R-SLAM（授權：CC BY-NC-SA 4.0）。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。