Adaptive 3DGS-SLAM (indoor digital twinning)
本研究針對室內建成環境的線上幾何數位孿生更新，以 MonoGS 預先建立的基準 3DGS 模型為先驗，提出自適應 3DGS-SLAM：新進 RGB-D 影格先以渲染比對方式對齊基準模型求位姿，再以高斯模糊後的滑動視窗 SSIM 產生變化遮罩，變化像素比例超過 2% 且符合原關鍵影格準則者才成為更新用關鍵影格；接著在 CUDA 反向傳播中累計各高斯對變化像素的貢獻，將高貢獻高斯及其鄰域隨機軟剪除一半，並只由遮罩區的 RGB-D 像素加入新高斯。ReplicaCAD 模擬中，相較 MonoGS 直接更新，每影格平均時間由 2.730 秒降至 1.162 秒（降低 57.4%），測試視角 SSIM 由 0.6537 升至 0.7542，ATE RMSE 約 0.105 m；在約 20 平方公尺實驗室以 Intel RealSense D435 實測，每影格時間降低 21.5%，SSIM 由 0.80 升至 0.84。全文未評估點雲或表面幾何精度，也沒有閉環與光束法平差。
本頁內容
RGB-D 3DGS-SLAM that tracks new frames against a prior MonoGS Gaussian model, detects changes with a sliding-window SSIM mask and edits only the Gaussians contributing to changed pixels; 57.4% faster and higher test-view SSIM than MonoGS updating on ReplicaCAD, 21.5% faster in a 20 m2 lab; rendering metrics only, ATE about 0.1 m, 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 | RGB-D camera: Intel RealSense D435, 1280 x 720 px at 30 frames per second (real-world case)、simulated RGB-D frames from ReplicaCAD (computer experiment) |
|---|---|
| 原文測試平台 | simulation (ReplicaCAD FRL apartment, simulated human roaming trajectories)、原文未報告 (real case: RealSense D435 in an indoor lab; handheld or robot carrier not stated) |
| 狀態估計 | frame-to-model tracking inherited from MonoGS: each frame's 6-DoF pose (unit quaternion and translation), initialised from the previous pose, is optimised by minimising a rendering-based loss between the captured image and the baseline 3DGS rendering (Eq. 4: L1 plus SSIM terms; Sec. 4.6 states RGB-D gives joint photometric and geometric supervision); no backend bundle adjustment |
| 資料關聯 | dense rendering-based alignment (no feature matching); change detection by sliding-window SSIM between Gaussian-blurred rendered and captured images, pixels below a threshold form a change mask, and a frame becomes an update keyframe only if the masked ratio exceeds 2% and the MonoGS keyframe criteria are also met |
| 時間表示 | discrete frame-by-frame processing of streaming RGB-D frames (30 fps capture in the real case); the design runs the frontend (tracking, change detection) in parallel with the backend (editing, mapping) as in MonoGS (Sec. 3.1), but all reported Python timings were run in single-thread mode (Sec. 4.3) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none (authors state the framework has no loop closure detection) |
| 全域最佳化 | none (no global or local bundle adjustment); only windowed keyframe 3DGS map optimisation as in MonoGS plus a final colour refinement of spherical harmonic coefficients |
| 地圖表示 | 3D Gaussians (MonoGS map); each Gaussian's contribution to changed pixels (alpha times transmittance, accumulated in the modified CUDA backward pass) selects candidates; candidates and neighbours within radius r are soft-pruned at random with probability 0.5 (opacity set to 0.01); new Gaussians are added only from masked RGB-D pixels and are protected while their source keyframe is in the window buffer |
| 先驗資訊 | baseline 3DGS model built beforehand with MonoGS from an earlier RGB-D session (ReplicaCAD Baked_sc0_staging_01; 1225 real frames in the lab case) |
| 可輸出幾何 | updated 3DGS model (about 128,251 Gaussians in the ReplicaCAD run, about 70% kept from the baseline); no point cloud or surface accuracy is evaluated; outputs assessed by rendering metrics (PSNR, SSIM) and ATE RMSE on simulation |
| 計算需求 | Python implementation on MonoGS with a modified CUDA rasteriser backward pass; AMD EPYC 9654 (16 cores in a VM), 60 GB RAM, single Nvidia 4090 GPU, Ubuntu 20.04, CUDA 11.6, single-thread mode; 1.162 s per frame on ReplicaCAD and 0.625 s per frame in the real case |
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 | Intel RealSense D435 | 方法輸入 | 未標示 | RGB-D images at 30 frames per second, 1280 x 720 px; 1652 frames (1225 for the baseline model, 427 for updating) | (Yuan et al., 2026, Sec. 5.1) |
| RGB-D 相機 | simulated RGB-D camera (human roaming trajectories; sensor not named) | 資料集感測器 | ReplicaCAD (FRL apartment, Baked_sc0_staging_01 and Baked_sc0_staging_05) | RGB-D sequences organised in the Replica dataset format; different paths for the baseline and update configurations; intrinsics not reported | (Yuan et al., 2026, Sec. 4.1) |
| 運算硬體 | AMD EPYC 9654 96-Core Processor | 執行運算平台 | 未標示 | 16 cores allocated within a virtual machine, 60 GB memory, Ubuntu 20.04 | (Yuan et al., 2026, Sec. 4.3, Sec. 5.2) |
| 運算硬體 | Nvidia 4090 GPU | 執行運算平台 | 未標示 | single GPU, CUDA 11.6 | (Yuan et al., 2026, Sec. 4.3, Sec. 5.2) |
作者報告的優勢與限制
優勢
- On ReplicaCAD, test-view SSIM 0.7542 and PSNR 19.48 dB versus 0.6537 and 13.98 dB for MonoGS updating (Table 3)
- Average time per frame 1.162 s versus 2.730 s for MonoGS (57.4% lower); keyframe mapping 2.772 s versus 9.345 s (Table 4)
- Real lab case: PSNR 23.37 dB, SSIM 0.84, 0.625 s per frame versus 22.66 dB, 0.80, 0.796 s for MonoGS updating; 37% fewer processed frames (Table 5, Sec. 5.4)
- Keeps unchanged regions that rebuild baselines (SplaTAM, Photo-SLAM, GS-ICP-SLAM) lose when only the update scan is used (Sec. 4.6, Fig. 8)
- Mitigates catastrophic forgetting and visual tearing seen in MonoGS updating (Fig. 12)
限制
- Tracking ATE RMSE about 0.105 m on ReplicaCAD, far worse than SplaTAM (0.0086 m) and Photo-SLAM (0.040 m), because of lightweight frame-to-model alignment without backend optimisation (Table 3, Sec. 4.6)
- No loop closure or bundle adjustment (Sec. 6.2)
- Assumes static environments; dynamic objects and occlusions not handled (Sec. 6.2)
- Real-world validation limited to a single laboratory case of about 20 m2 (Sec. 5.1, Sec. 6.2)
- No semantic information (Sec. 6.2)
- RGB-only configuration suffers depth ambiguity: ATE RMSE 1.4795 m and 2.321 s per frame (Table 3, Sec. 4.6)
- No point or surface geometric accuracy metric is reported anywhere in the full text (Sec. 4 to 5)
營建工程相關證據
營建資訊期刊中處理「既有 3DGS 數位孿生於例行巡檢時增量更新」的研究（作者指出 3DGS-SLAM 的增量更新仍少有探討），對設施管理與室內改修追蹤有參考價值；但驗證僅有 ReplicaCAD 模擬與單一約 20 平方公尺實驗室，評估以渲染指標（PSNR、SSIM）與模擬 ATE 為主，未量測幾何誤差，模擬中的追蹤誤差約 10 公分且無閉環，不宜作為量測等級竣工幾何的證據。4D 施工進度監測只是作者在討論中的展望，並未經實驗驗證。
原文驗證環境：模擬、受控實驗
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 43 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。
Yuan et al., 2026 · Table 2 本方法 24 筆
資料集與序列ReplicaCAD (FRL apartment) · update Baked_sc0_staging_05 (baseline from Baked_sc0_staging_01)
表格設定（擷取紀錄原文）：Change-ratio threshold sensitivity of the proposed method on ReplicaCAD; training-view metrics on update frames, test-view metrics on 100 separately generated poses; Avg. time = total time / scanned frames (Yuan et al., 2026, Table 2)
Training-view SSIM，ReplicaCAD (FRL apartment) · update Baked_sc0_staging_05 (baseline from Baked_sc0_staging_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 Yuan et al., 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(Yuan et al., 2026, Table 2)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Proposed method, change ratio threshold >=1%本方法原文提出 | 0.8295 | (Yuan et al., 2026, Table 2) |
| Proposed method, change ratio threshold >=2% (default)本方法原文提出 | 0.8354 | (Yuan et al., 2026, Table 2) |
| Proposed method, change ratio threshold >=3.5%本方法原文提出 | 0.8329 | (Yuan et al., 2026, Table 2) |
| Proposed method, change ratio threshold >=5%本方法原文提出 | 0.824 | (Yuan et al., 2026, Table 2) |
Yuan et al., 2026 · Table 3 本方法 8 筆
資料集與序列ReplicaCAD (FRL apartment) · update Baked_sc0_staging_05 (1000 frames)
表格設定（擷取紀錄原文）：ReplicaCAD comparison: proposed and MonoGS update the baseline model, SplaTAM, Photo-SLAM and GS-ICP-SLAM rebuild from the update sequence only; test views = 100 generated poses; Photo-SLAM and GS-ICP-SLAM times from default multi-process runs, others single-thread Python (Yuan et al., 2026, Table 3)
Test-view SSIM，ReplicaCAD (FRL apartment) · update Baked_sc0_staging_05 (1000 frames)
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 Yuan et al., 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(Yuan et al., 2026, Table 3)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Proposed method本方法原文提出 | 0.7542 | (Yuan et al., 2026, Table 3) |
| Update with MonoGS | 0.6537 | (Yuan et al., 2026, Table 3) |
| Rebuild with SplaTAM | 0.6834 | (Yuan et al., 2026, Table 3) |
| Rebuild with Photo-SLAM | 0.7339 | (Yuan et al., 2026, Table 3) |
| Rebuild with GS-ICP-SLAM | 0.581 | (Yuan et al., 2026, Table 3) |
| Proposed method (RGB-only)本方法原文提出 | 0.6696 | (Yuan et al., 2026, Table 3) |
Yuan et al., 2026 · Table 4 本方法 6 筆
資料集與序列ReplicaCAD (FRL apartment) · update Baked_sc0_staging_05 (1000 frames)
表格設定（擷取紀錄原文）：Per-step runtime on the ReplicaCAD update sequence (1000 frames); tracking and keyframe determination averaged over all frames, Gaussian edit and mapping over keyframes only; whole-process average = total time / scanned frames; single-thread Python (Yuan et al., 2026, Table 4)
Whole process Avg. time [s]，ReplicaCAD (FRL apartment) · update Baked_sc0_staging_05 (1000 frames)
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 Yuan et al., 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(Yuan et al., 2026, Table 4)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Proposed method本方法原文提出硬體：AMD EPYC 9654 (16 cores in a VM), 60 GB RAM, single Nvidia 4090 GPU, Ubuntu 20.04, CUDA 11.6; single-thread Python | 1.162 s | (Yuan et al., 2026, Table 4) |
| Update with MonoGS硬體：AMD EPYC 9654 (16 cores in a VM), 60 GB RAM, single Nvidia 4090 GPU, Ubuntu 20.04, CUDA 11.6; single-thread Python | 2.73 s | (Yuan et al., 2026, Table 4) |
| Rebuild with SplaTAM硬體：AMD EPYC 9654 (16 cores in a VM), 60 GB RAM, single Nvidia 4090 GPU, Ubuntu 20.04, CUDA 11.6; single-thread Python | 5.988 s | (Yuan et al., 2026, Table 4) |
Yuan et al., 2026 · Table 5 本方法 3 筆
資料集與序列authors' real-world RGB-D dataset · PSNR and SSIM: 300 random frames evenly sampled from the initial (1225) and updating (427) subsets; Avg. time: total runtime divided by the number of frames in the dataset (Sec. 5.4)
表格設定（擷取紀錄原文）：Real-world lab case (Intel RealSense D435; 1225 baseline frames then 427 update frames); PSNR and SSIM are means over 300 random frames sampled from both subsets; Avg. time = total runtime / all frames (Yuan et al., 2026, Table 5)
PSNR [dB]，authors' real-world RGB-D dataset · PSNR and SSIM: 300 random frames evenly sampled from the initial (1225) and updating (427) subsets; Avg. time: total runtime divided by the number of frames in the dataset (Sec. 5.4)
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 Yuan et al., 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(Yuan et al., 2026, Table 5)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Adaptive 3DGS-SLAM本方法原文提出 | 23.37 dB | (Yuan et al., 2026, Table 5) |
| MonoGS updating pipeline | 22.66 dB | (Yuan et al., 2026, Table 5) |
| Baseline without updating (MonoGS initial model) | 21.27 dB | (Yuan et al., 2026, Table 5) |
其他比較組
來源
Yuan et al., 2026
（2026）Adaptive 3DGS-SLAM-driven incremental online geometric digital twinning complex indoor built environmentsAdvanced Engineering Informatics, 76, 105077
同儕審查已出版已讀全文近十年查證後修正