MCGS SLAM
MCGS-SLAM 以 MonoGS 高斯潑濺 SLAM 為基礎,加入由位姿歷史或加速度計推得的運動先驗(含自適應權重與快速運動偵測)、依位置穩定性、形狀、不透明度、空間與邊緣重要性計算的高斯信心度、依速度、旋轉、位移、可見重疊與視覺豐富度的自適應關鍵影格選擇,以及信心加權的多任務最佳化。評估只用 TUM RGB-D、Replica 與 EuRoC 公開資料集;TUM fr1_desk 的 ATE RMSE 由 2.13 cm 降至 1.52 cm,但 Replica 七個場景有三個變差,EuRoC MH03 至 MH05 誤差仍達 1.79 至 3.78 m,且未在實際建物或工地測試。
本頁內容
MonoGS-based RGB-D 3DGS SLAM with a motion prior, per-Gaussian confidence and adaptive keyframing; evaluated only on TUM, Replica and EuRoC benchmarks, not on building data.
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 (synchronized RGB-D input)、optional accelerometer and gyroscope for the motion prior; a motion-only mode infers velocity and acceleration from pose history when inertial data are absent |
|---|---|
| 原文測試平台 | 未記錄 |
| 狀態估計 | frame-to-model tracking by minimizing photometric and geometric rendering losses on a 3DGS map (MonoGS baseline) plus an adaptively weighted motion-prior loss (translation, rotation, velocity and acceleration smoothness), with fast-motion detection that raises the prior weight and keyframe rate |
| 資料關聯 | dense direct photometric and geometric residuals over visible Gaussians, weighted by per-Gaussian confidence (position stability, shape quality, opacity, spatial importance, edge importance); keyframes chosen by velocity, rotation, displacement, visibility overlap and feature richness |
| 時間表示 | discrete per-frame poses with a velocity and acceleration motion prior from a configurable pose-history window (or IMU integration when available) |
| 去畸變 | 不適用 (camera frames; no scanning sensor) |
| 迴圈閉合 | none (authors state the confidence-aware map maintenance avoids descriptor-based loop closure; EuRoC table lists MCGS-SLAM without loop closure) |
| 全域最佳化 | confidence-weighted bundle adjustment over keyframes in a sliding window during map maintenance; no global pose graph |
| 地圖表示 | 3D Gaussian splatting map (MonoGS baseline) with per-Gaussian confidence driving pruning, densification and reduced learning rates for stable Gaussians |
| 先驗資訊 | none beyond the motion prior (pose history or IMU); no BIM or scene prior |
| 可輸出幾何 | 3D Gaussian map with rendered RGB and depth and a reconstructed point cloud; geometry assessed qualitatively and by cloud-to-mesh distance ranges on Replica (Fig. 15), no tabulated geometric accuracy |
| 計算需求 | NVIDIA RTX 4090 (paper states 48 GB VRAM) + Intel Core i9-13900K; 1.09 to 1.84 FPS on Replica, 1.08 FPS for the full system on TUM fr1_desk; 4.63 GB GPU memory; tracking 15.85 ms and mapping 88 ms per iteration |
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 |
|---|---|---|---|---|---|
| RGB-D 相機 | Microsoft Kinect | 資料集感測器 | TUM RGB-D | sensor of the TUM real-world indoor sequences | (Deng & Gan, 2026, Sec. 4.1) |
| 運算硬體 | NVIDIA RTX 4090 | 執行運算平台 | 未標示 | stated as 48 GB VRAM (Sec. 4.1); MonoGS reproduced on it (Table 8 note); Orbeez-SLAM, Point-SLAM and SplaTAM FPS cited from GS-ICP-SLAM, which also used an RTX 4090; hardware for the Table 9 baseline memory values not stated | (Deng & Gan, 2026, Sec. 4.1; Table 8 note) |
| 運算硬體 | Intel Core i9-13900K | 執行運算平台 | 未標示 | desktop CPU | (Deng & Gan, 2026, Sec. 4.1) |
| 其他 | motion capture system (model not reported) | 參考或真值量測 | TUM RGB-D | ground-truth trajectories of TUM | (Deng & Gan, 2026, Sec. 4.1) |
作者報告的優勢與限制
優勢
- TUM fr1_desk ATE RMSE 1.52 cm versus 2.13 cm for baseline MonoGS (Table 1)
- TUM fr3_walking_xyz ATE RMSE 0.25 m versus 0.73 m for MonoGS (Table 2)
- Replica average ATE improvement 11.03% (up to 83.09% in room1); PSNR +2.03%, SSIM +0.63%, LPIPS 11.76% (Tables 3 and 6)
- With 25% of frames on fr1_desk, ATE RMSE 2.67 cm versus 6.64 cm for MonoGS (Table 7)
限制
- Evaluation limited to benchmark datasets; authors plan to collect a new dataset (Sec. 5)
- Replica ATE worse than MonoGS in 3 of 7 scenes (office2, office3, room0) (Table 3)
- EuRoC MH03 to MH05 ATE 1.79 to 3.78 m, far above ORB-SLAM3 stereo (0.02 to 0.09 m); MH05 worse than MonoGS (3.78 versus 3.19 m) (Table 11)
- Full system is slower and heavier than MonoGS on fr1_desk (1.08 versus 1.98 FPS, 4.63 versus 3.21 GB) and its ATE (1.52 cm) is worse than the motion-prior-only (1.46 cm) or confidence-only (1.43 cm) variants (Table 10)
- Much less accurate than dynamic-scene methods such as DynaSLAM (0.02 m) on fr3_walking_xyz (Table 2)
- Many thresholds and weights are configuration parameters that the authors say can be tuned per dataset or sensor setup (Sec. 3.2 to 3.4; Sec. 5)
營建工程相關證據
題名與動機指向室內建成環境與 Scan-to-BIM,但實驗只用 TUM RGB-D、Replica 合成資料與 EuRoC 大型室內序列,沒有建物或工地資料;作者在結論承認評估僅限基準資料集。可作為 3DGS SLAM 在弱紋理與快速運動情境的方法參考,不能作為營建場域精度的證據。
原文驗證環境:模擬、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 6 個比較組,合計 24 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 2 組列在最後,並連到性能比較頁。
Deng & Gan, 2026 · Table 3 本方法 7 筆
指標ATE RMSE (cm)
表格設定(擷取紀錄原文):Replica pose tracking, baseline MonoGS versus MCGS-SLAM (Deng & Gan, 2026, Table 3)
ATE RMSE (cm),Replica · 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 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。
資料來源作者報告值(Deng & Gan, 2026, Table 3)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Baseline MonoGS | 0.444 cm | (Deng & Gan, 2026, Table 3) |
| MCGS-SLAM (Ours)本方法原文提出 | 0.42 cm | (Deng & Gan, 2026, Table 3) |
Deng & Gan, 2026 · Table 8 本方法 7 筆
指標FPS
表格設定(擷取紀錄原文):Replica FPS; MonoGS reproduced on the same RTX 4090 (Orbeez-SLAM, Point-SLAM and SplaTAM columns cited from GS-ICP-SLAM, not transcribed) (Deng & Gan, 2026, Table 8)
FPS,Replica · O0
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 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。
資料來源作者報告值(Deng & Gan, 2026, Table 8)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MonoGS (re-run by authors)硬體:NVIDIA RTX 4090 + Intel Core i9-13900K | 1.5 Hz | (Deng & Gan, 2026, Table 8) |
| MCGS-SLAM (Ours)本方法原文提出硬體:NVIDIA RTX 4090 + Intel Core i9-13900K | 1.5 Hz | (Deng & Gan, 2026, Table 8) |
Deng & Gan, 2026 · Table 11 本方法 5 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。
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 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。
資料來源作者報告值(Deng & Gan, 2026, Table 11)
| 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) | 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) | 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)本方法原文提出 | 0.05 m | (Deng & Gan, 2026, Table 11) |
Deng & Gan, 2026 · Table 1 本方法 3 筆
指標ATE RMSE (cm)
表格設定(擷取紀錄原文):TUM static scenes; ATE RMSE; the paper does not state whether baseline values were re-run or taken from the literature (Deng & Gan, 2026, Table 1)
ATE RMSE (cm),TUM RGB-D · freiburg1_desk
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 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。
資料來源作者報告值(Deng & Gan, 2026, Table 1)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| SplaTAM | 3.35 cm | (Deng & Gan, 2026, Table 1) |
| MM3DGS-SLAM | 3.51 cm | (Deng & Gan, 2026, Table 1) |
| ORB-SLAM3 | 2.07 cm | (Deng & Gan, 2026, Table 1) |
| CVO-SLAM | 3.15 cm | (Deng & Gan, 2026, Table 1) |
| RKD-SLAM | 2.1 cm | (Deng & Gan, 2026, Table 1) |
| CG-SLAM | 2.4 cm | (Deng & Gan, 2026, Table 1) |
| GS-ICP-SLAM | 2.7 cm | (Deng & Gan, 2026, Table 1) |
| GLC-SLAM | 1.85 cm | (Deng & Gan, 2026, Table 1) |
| NICE-SLAM | 4.3 cm | (Deng & Gan, 2026, Table 1) |
| Uni-SLAM | 2.37 cm | (Deng & Gan, 2026, Table 1) |
| DI-Fusion (reference [57] is DI-SLAM) | 4.4 cm | (Deng & Gan, 2026, Table 1) |
| Vox-Fusion | 3.52 cm | (Deng & Gan, 2026, Table 1) |
| ESLAM | 2.47 cm | (Deng & Gan, 2026, Table 1) |
| Co-SLAM | 2.4 cm | (Deng & Gan, 2026, Table 1) |
| Gaussian-SLAM | 2.73 cm | (Deng & Gan, 2026, Table 1) |
| Point-SLAM | 4.3 cm | (Deng & Gan, 2026, Table 1) |
| Baseline MonoGS | 2.13 cm | (Deng & Gan, 2026, Table 1) |
| MCGS-SLAM (Ours)本方法原文提出 | 1.52 cm | (Deng & Gan, 2026, Table 1) |
其他比較組
來源
Deng & Gan, 2026
(2026)Motion-prior and Confidence-aware Gaussian Splatting (MCGS) SLAM for 3D scene reconstruction of indoor built environmentsAdvanced Engineering Informatics, 71, 104317
同儕審查已出版已讀全文近十年查證後修正