SLAM 點雲圖鑑Point clouds for construction
M-01 · 2015

ElasticFusion

ElasticFusion 以面元（surfel）表示稠密地圖，採用由目前影像對模型（frame-to-model）的稠密追蹤與時間視窗內的面元融合。系統盡量頻繁地做局部模型對模型迴圈閉合，並以隨機蕨（randomised fern）影像編碼偵測全域迴圈，再以非剛性變形直接校正地圖，而不使用位姿圖或事後處理。作者將適用範圍定為房間尺度。

具全域修正的完整 SLAM經典已讀全文同儕審查已出版視覺、視覺慣性與 RGB-D 稠密 SLAM

閱讀版本
RSS XI (2015) proceedings PDF, version of record, DOI 10.15607/RSS.2015.XI.001 (9 pages); IJRR 2016 extension not read
全文取得途徑
出版者開放取用（open access）
本頁內容

ElasticFusion maintains a surfel map with dense frame-to-model tracking and applies frequent local and fern-based global loop closures as non-rigid map deformations instead of pose-graph optimisation, at room scale.

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handheld
狀態估計Gauss-Newton on E_track = E_icp + 0.1 E_rgb over a three-level coarse-to-fine pyramid (6x6 normal equations via CUDA tree reduction, Cholesky on CPU); loop closures applied through an embedded deformation graph rebuilt each frame by systematic sampling of surfels, nodes connected by initialisation time (k = 4), optimised by Gauss-Newton with sparse Cholesky on the CPU using rotation, regularisation, constraint and pin terms (weights 1, 10, 100, 100)
資料關聯frame-to-model: point-to-plane ICP with projective data association between the live depth map and the splatted active-model depth prediction, plus photometric intensity error between the live colour image and the splatted model colour prediction (weight 0.1); the same registration is used for model-to-model loop alignment
時間表示discrete poses; surfels carry timestamps for time-windowed fusion (abstract)
去畸變原文未報告
迴圈閉合local: each frame the active model prediction is registered to the inactive model prediction and accepted if residual, inlier count and covariance eigenvalue checks pass, then the map is deformed and the region reactivated; global: randomised fern database of predicted views at 80x60, matched views registered and accepted only if the resulting deformation is consistent with the map geometry; the fern database can also serve relocalisation, not needed in the evaluated data
全域最佳化non-rigid deformation of the surfel map instead of pose-graph optimisation (abstract; conclusion)
地圖表示unordered list of surfels (position, normal, colour, weight, radius, initialisation and last-update timestamps) split into active and inactive sets by a time window; up to 4.8 million surfels in the qualitative scans
先驗資訊none
可輸出幾何surfel map (oriented points with radius and colour)
計算需求desktop PC with Intel Core i7-4930K 3.4 GHz, 32 GB RAM and nVidia GeForce GTX 780 Ti with 3 GB; CUDA for tracking reduction and OpenGL shading language for prediction and map management; frame time rises with surfel count, overall average 31 ms and peak average 45 ms (worst case 22 Hz) on the Hotel sequence

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standard RGB-D camera (Microsoft Kinect or ASUS Xtion Pro Live named as examples)方法輸入ElasticFusion qualitative datasets (Copy from Zhou and Koltun, Lab, Hotel, Office)device used for the authors' hand-held qualitative datasets not stated(Whelan et al., 2015a, Sec. II footnote 1; Sec. VII-B)
運算硬體Intel Core i7-4930K執行運算平台未標示3.4GHz, 32GB of RAM(Whelan et al., 2015a, Sec. VII-C)
運算硬體nVidia GeForce GTX 780 Ti執行運算平台未標示3GB of memory(Whelan et al., 2015a, Sec. VII-C)
其他highly precise motion capture system參考或真值量測TUM RGB-Dsynchronised ground-truth poses of the TUM RGB-D benchmark(Whelan et al., 2015a, Sec. VII-A)

作者報告的優勢與限制

優勢

  • Globally consistent room-scale surfel maps online without pose graph or post-processing (abstract)
  • Frequent non-rigid deformations improved both trajectory and surface reconstruction in authors' evaluation (conclusion; Tables I to III)
  • Lowest surface reconstruction error on all four ICL-NUIM living-room sequences (Table III)
  • Frame-to-model tracking alone is already comparable to pose-graph systems on TUM RGB-D (Sec. VII-A)

限制

  • Designed and evaluated for room-scale environments; scalability beyond whole rooms left to future work (conclusion)
  • Relies on GPU programming for tracking, prediction and map management (Sec. II)
  • Frame processing time grows with the number of surfels in the map (Sec. VII-C; Fig. 6)

營建工程相關證據

論文未報告營建測試；表面精度評估使用 ICL-NUIM 合成資料。房間尺度與 GPU 需求限制其直接用於整層或整棟建築（推論）。

原文驗證環境：模擬、公開基準、獨立參考量測

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 221 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。

Schöps et al., 2020 · Table 2 (ground-truth trajectories) 本方法 24 筆

表格設定（擷取紀錄原文）：ICL-NUIM living room with simulated depth noise; ground-truth trajectories used and loop-closure handling disabled for all methods; reconstructions aligned to the ground-truth model with point-to-plane ICP; evaluation threshold 1 cm; 'smoothed' = same bilateral filter as SurfelMeshing preprocessing (Schöps et al., 2020, Table 2 (ground-truth trajectories))

Accuracy [%] (share of reconstructed surfels within 1 cm of ground truth)，ICL-NUIM · 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。

這些是 Schöps et al., 2020 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：原文未報告；對齊方式：SE(3) 剛體對齊；單位：%；場景：synthetic indoor living room

資料來源作者報告值(Schöps et al., 2020, Table 2 (ground-truth trajectories))

數值與出處
方法（原文寫法）報告值出處
InfiniTAM [29]76.4%(Schöps et al., 2020, Table 2)
InfiniTAM [29] - smoothed78.3%(Schöps et al., 2020, Table 2)
FastFusion [27]85.5%(Schöps et al., 2020, Table 2)
FastFusion [27] - smoothed75.9%(Schöps et al., 2020, Table 2)
ElasticFusion [17]本方法96.2%(Schöps et al., 2020, Table 2)
ElasticFusion [17] - smoothed本方法95.7%(Schöps et al., 2020, Table 2)
SurfelMeshing (Ours)原文提出93.5%(Schöps et al., 2020, Table 2)

Scona et al., 2018 · Table I 本方法 22 筆

表格設定（擷取紀錄原文）：TUM (Freiburg) RGB-D sequences grouped as static (fr1), low dynamic (fr3/sit) and high dynamic (fr3/walk) environments; StaticFusion and VO-SF at QVGA, ElasticFusion and Co-Fusion at their default VGA; fr3/walk_halfsphere* skips the first 5 s of high dynamics; relative pose error per second (Scona et al., 2018, Table I)

Trans. RPE RMSE (cm/s)，TUM RGB-D (Freiburg) · fr1/xyz

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）

這些是 Scona et al., 2018 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/s；場景：indoor office scenes, static to highly dynamic (walking people), RGB-D camera (carrying mode not stated in the paper)

資料來源作者報告值(Scona et al., 2018, Table I)

數值與出處
方法（原文寫法）報告值出處
VO-SF (Jaimez et al. joint visual odometry and scene flow)2.1 cm/s(Scona et al., 2018, Table I)
EF (ElasticFusion)本方法1.9 cm/s(Scona et al., 2018, Table I)
CF (Co-Fusion)2.3 cm/s(Scona et al., 2018, Table I)
BaMVO (Kim et al.)無數值未報告註記（擷取紀錄）：not reported (BaMVO shown only for sequences evaluated in its original publication)(Scona et al., 2018, Table I)
SF (StaticFusion)原文提出2.3 cm/s(Scona et al., 2018, Table I)

Schöps et al., 2020 · Table 2 (loop-closure trajectories) 本方法 18 筆

表格設定（擷取紀錄原文）：ICL-NUIM living room with simulated depth noise; trajectories estimated with ElasticFusion including loop closures (kt3 omitted because ElasticFusion failed); aligned with point-to-plane ICP; threshold 1 cm; InfiniTAM and FastFusion cannot handle loop closures and have no values (Schöps et al., 2020, Table 2 (loop-closure trajectories))

Accuracy [%] (share of reconstructed surfels within 1 cm of ground truth)，ICL-NUIM · 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。

這些是 Schöps et al., 2020 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：原文未報告；對齊方式：SE(3) 剛體對齊；單位：%；場景：synthetic indoor living room

資料來源作者報告值(Schöps et al., 2020, Table 2 (loop-closure trajectories))

數值與出處
方法（原文寫法）報告值出處
ElasticFusion [17]本方法95.8%(Schöps et al., 2020, Table 2)
ElasticFusion [17] - smoothed本方法96.8%(Schöps et al., 2020, Table 2)
SurfelMeshing (Ours)原文提出87.2%(Schöps et al., 2020, Table 2)

Scona et al., 2018 · Table II 本方法 11 筆

指標Trans. ATE RMSE (cm)

表格設定（擷取紀錄原文）：TUM (Freiburg) RGB-D sequences grouped as static (fr1), low dynamic (fr3/sit) and high dynamic (fr3/walk) environments; StaticFusion and VO-SF at QVGA, ElasticFusion and Co-Fusion at their default VGA; fr3/walk_halfsphere* skips the first 5 s of high dynamics (Scona et al., 2018, Table II)

Trans. ATE RMSE (cm)，TUM RGB-D (Freiburg) · fr1/xyz

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 Scona et al., 2018 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：indoor office scenes, static to highly dynamic (walking people), RGB-D camera (carrying mode not stated in the paper)

資料來源作者報告值(Scona et al., 2018, Table II)

數值與出處
方法（原文寫法）報告值出處
VO-SF (Jaimez et al. joint visual odometry and scene flow)5.1 cm(Scona et al., 2018, Table II)
EF (ElasticFusion)本方法1.2 cm(Scona et al., 2018, Table II)
CF (Co-Fusion)1.4 cm(Scona et al., 2018, Table II)
SF (StaticFusion)原文提出1.4 cm(Scona et al., 2018, Table II)

其他比較組

列出其餘 32 個比較組

來源

  • Whelan et al., 2015a

    Thomas Whelan, Stefan Leutenegger, Renato Salas Moreno, Ben Glocker, Andrew Davison（2015）ElasticFusion: Dense SLAM Without A Pose GraphRobotics: Science and Systems XI

    DOI 10.15607/rss.2015.xi.001程式碼

    同儕審查已出版已讀全文經典

程式碼：https://github.com/mp3guy/ElasticFusion（授權：custom licence, non-commercial, internal or academic research purposes only (LICENSE.txt)）。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