SLAM 點雲圖鑑Point clouds for construction
M-01 · 2023

Point-SLAM

Point-SLAM 將神經特徵錨定在隨輸入逐步生成的點雲上，並依影像梯度動態調整點密度，細節處加密、平坦處稀疏；追蹤與建圖共用同一個以 RGB-D 重渲染誤差最佳化的點式表示。其網格評估在計算精確率與召回率前先以 ICP 對齊，因此量到的是局部形狀品質而非全域位置精度。

里程計＋局部建圖近十年已讀全文同儕審查已出版學習式、神經隱式與高斯潑濺 SLAM

閱讀版本
arXiv v3 (2023-09-12) read in full including supplementary App. A to E; CVF open-access ICCV 2023 main paper also read: Tables 1 to 4, Table 6, Fig. 3a values, hardware and limitation text identical to arXiv v3
全文取得途徑
arXiv 預印本
本頁內容

Anchors neural features on an input-adaptive point cloud used for both tracking and mapping in RGB-D SLAM.

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
原文測試平台未記錄
狀態估計gradient-based minimization of an RGB-D re-rendering loss for tracking and mapping, run as separate alternating processes; tracking initialised with a constant-speed assumption; mapping iterations adapted to the number of newly added points; optional exposure-compensation MLP for ScanNet
資料關聯direct colour and depth re-rendering loss
時間表示discrete poses
去畸變不適用
迴圈閉合none (authors note a gap to traditional methods with loop closures)
全域最佳化none
地圖表示neural point cloud: features anchored at points with density adapted to image-gradient information
先驗資訊uses the pretrained and fixed middle-level geometric decoder provided by NICE-SLAM; colour decoder, interpolation MLP and point features are optimized online
可輸出幾何mesh produced by rendering depth and colour every fifth frame along the estimated trajectory, TSDF Fusion at 1 cm voxels and marching cubes; evaluated with precision, recall and F-score at a 1 cm threshold after ICP alignment to the GT mesh; rendered depth and colour images
計算需求runtime profiled on a single NVIDIA RTX 2080 Ti; experiments on various NVIDIA GPUs with at most 12 GB memory (Sec. 4.4; App. C); on Replica office 0: 21 ms tracking and 33 ms mapping per iteration, 0.85 s tracking and 9.85 s mapping per frame, 27.23 MB scene embedding (Table 6); （Huang et al., 2024c） Table 1 measured 0.345 tracking FPS and more than 2 h operation time on Replica RGB-D with an RTX 4090

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NVIDIA RTX 2080 Ti執行運算平台未標示single GPU used to profile Point-SLAM and NICE-SLAM runtimes(Sandström et al., 2023, Sec. 4.4 Memory and Runtime Analysis)
運算硬體NVIDIA RTX 3090執行運算平台未標示GPU used for the Vox-Fusion runtime in Table 6(Sandström et al., 2023, Sec. 4.4)
其他external motion capture system (model not named)參考或真值量測TUM-RGBDprovides TUM-RGBD ground-truth poses(Sandström et al., 2023, Sec. 4 Datasets)

作者報告的優勢與限制

優勢

  • Better reconstruction and rendering accuracy than NICE-SLAM, Vox-Fusion and ESLAM on Replica (Sec. 4.1)

限制

  • More sensitive to motion blur and specularities (Sec. 4.2; Limitations)
  • Point density follows a heuristic; many empirical hyperparameters (Limitations)
  • No loop closure; gap to traditional methods on real data (Sec. 4.2)
  • Point locations are not optimized on the fly, limiting robustness to depth noise (Limitations)
  • The non-linear appearance MLP is disabled on TUM-RGBD and ScanNet because it does not help when tracking errors are higher (Sec. 4.4)
  • (derived from Table 6) about 0.85 s tracking and 9.85 s mapping per frame on Replica office 0

營建工程相關證據

論文未涉及營建場域；資料為 Replica、TUM-RGBD、ScanNet。

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 36 個比較組，合計 169 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 32 組列在最後，並連到性能比較頁。

Keetha et al., 2024 · Table 1 本方法 25 筆

指標ATE RMSE [cm]

表格設定（擷取紀錄原文）：Online camera-pose estimation, ATE RMSE [cm]; Baseline numbers taken from Point-SLAM; SplaTAM averaged over 3 seeds (Keetha et al., 2024, Table 1)

ATE RMSE [cm]，TUM-RGBD · 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。

這些是 Keetha et al., 2024 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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：cm；場景：real RGB-D sequences from old low-quality cameras (sparse depth, strong motion blur)

資料來源作者報告值(Keetha et al., 2024, Table 1)

數值與出處
方法（原文寫法）報告值出處
Point-SLAM本方法8.92 cm(Keetha et al., 2024, Table 1)
SplaTAM原文提出5.48 cm(Keetha et al., 2024, Table 1)
Vox-Fusion11.31 cm(Keetha et al., 2024, Table 1)
NICE-SLAM15.87 cm(Keetha et al., 2024, Table 1)
Kintinuous4.84 cm(Keetha et al., 2024, Table 1)
ElasticFusion6.91 cm(Keetha et al., 2024, Table 1)
ORB-SLAM21.98 cm(Keetha et al., 2024, Table 1)

Ha et al., 2024 · Table 1 本方法 9 筆

指標ATE RMSE [cm]

表格設定（擷取紀錄原文，這些數值分屬表中不同部分）：(Ha et al., 2024, Table 1)

  • Replica ATE RMSE; * = reproduced with official code; GS-SLAM from its paper, Photo-SLAM only average from its paper
  • Replica ATE RMSE; * = reproduced with official code; GS-SLAM from its paper, Photo-SLAM only average from its paper; the Gaussian Splatting SLAM (MonoGS) row appears only in the ECCV version of record, reproduced with official code and evaluated on keyframes only
  • Replica ATE RMSE; the ECCV 2024 version of record (Table 1) adds an ORB-SLAM3 average taken from Photo-SLAM [12]; per-scene cells are '-'

ATE RMSE [cm]，Replica · average of 8 scenes

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。

這些是 Ha et al., 2024 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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：cm；場景：synthetic indoor scenes

資料來源作者報告值(Ha et al., 2024, Table 1)

數值與出處
方法（原文寫法）報告值出處
NICE-SLAM* [ 47 ]1.42 cm(Ha et al., 2024, Table 1)
Point-SLAM* [ 32 ]本方法0.54 cm(Ha et al., 2024, Table 1)
GS-SLAM [ 44 ]0.5 cm(Ha et al., 2024, Table 1)
Photo-SLAM [ 12 ]0.6 cm(Ha et al., 2024, Table 1)
SplaTAM* [ 14 ]0.36 cm(Ha et al., 2024, Table 1)
Ours (limited to 30 FPS)原文提出0.16 cm(Ha et al., 2024, Table 1)
Gaussian Splatting SLAM* [22] (ECCV version only)0.32 cm(Ha et al., 2024, Table 1 (ECCV 2024 version of record))
ORB-SLAM3 [4] (ECCV version only)1.8 cm(Ha et al., 2024, Table 1 (ECCV 2024 version of record))

Yan et al., 2024 · Table 1 本方法 9 筆

指標ATE RMSE [cm]

表格設定（擷取紀錄原文）：Replica ATE RMSE, 8 scenes; * = reproduced with official code; methods in upper part run below 5 FPS (Yan et al., 2024, Table 1)

ATE RMSE [cm]，Replica · Rm0

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。

這些是 Yan et al., 2024 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。

統計量：均方根誤差（RMSE）；對齊方式：原文未報告；單位：cm；場景：synthetic indoor scenes

資料來源作者報告值(Yan et al., 2024, Table 1)

數值與出處
方法（原文寫法）報告值出處
Point-SLAM [ 27 ]本方法0.56 cm(Yan et al., 2024, Table 1)
NICE-SLAM [ 55 ]0.97 cm(Yan et al., 2024, Table 1)
Vox-Fusion ∗ [ 48 ]1.37 cm(Yan et al., 2024, Table 1)
ESLAM [ 11 ]0.71 cm(Yan et al., 2024, Table 1)
CoSLAM [ 41 ]0.7 cm(Yan et al., 2024, Table 1)
Ours原文提出0.48 cm(Yan et al., 2024, Table 1)

Matsuki et al., 2024 · Table 2 本方法 9 筆

指標ATE RMSE (keyframes)

表格設定（擷取紀錄原文）：Keyframe ATE RMSE (cm) on Replica, RGB-D only (Replica has purely rotational motions); baselines from Point-SLAM; Ours = multi-process real-time implementation, Ours (sp) = single-process with more mapping iterations. (Matsuki et al., 2024, Table 2)

ATE RMSE (keyframes)，Replica · room0

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。

這些是 Matsuki et al., 2024 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。

統計量：均方根誤差（RMSE）；對齊方式：SE(3) 剛體對齊；單位：cm；場景：synthetic Replica sequences (room0 to room2, office0 to office4) with purely rotational camera motion, RGB-D input

資料來源作者報告值(Matsuki et al., 2024, Table 2)

數值與出處
方法（原文寫法）報告值出處
iMAP [35]3.12 cm(Matsuki et al., 2024, Table 2)
NICE-SLAM0.97 cm(Matsuki et al., 2024, Table 2)
Vox-Fusion [45]1.37 cm(Matsuki et al., 2024, Table 2)
ESLAM [9]0.71 cm(Matsuki et al., 2024, Table 2)
Point-SLAM [29]本方法0.61 cm(Matsuki et al., 2024, Table 2)
MonoGS (Ours, multi-process)原文提出0.44 cm(Matsuki et al., 2024, Table 2)
MonoGS (Ours sp, single-process)原文提出0.33 cm(Matsuki et al., 2024, Table 2)

其他比較組

列出其餘 32 個比較組

來源

  • Sandström et al., 2023

    Erik Sandström, Yue Li, Luc Van Gool, Martin R. Oswald（2023）Point-SLAM: Dense Neural Point Cloud-based SLAM2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 18387-18398

    DOI 10.1109/iccv51070.2023.01690arXiv 2304.04278程式碼

    同儕審查已出版已讀全文近十年

程式碼：https://github.com/eriksandstroem/Point-SLAM（授權：Apache-2.0）。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。

回到方法圖鑑

SLAM 點雲圖鑑

本站為綜述論文《SLAM-Based Point Clouds for Construction: A Review》研究過程的互動說明文件。互動圖為教學模擬，不代表任何特定系統的實測表現；方法描述、數值與結論以所引原始文獻為準。

內容基準日 2026-09-25 · 研究與網站製作使用 AI 輔助工具，見 證據標記與方法

選擇開啟Esc關閉