Dense Surfel Mapping (Wang, Gao, Shen)
作者提出只用 CPU 的面元（surfel）稠密建圖系統，相機位姿、參考關鍵影格與位姿圖都由外部稀疏視覺 SLAM（ORB-SLAM2 或 VINS-Mono）提供。每張影像先以擴充的 SLIC 依強度、深度與位置分割超像素，可處理無效深度，並以 Huber 損失求穩健的平均深度，再由超像素建立面元，因此能融合 RGB-D、立體相機或單眼深度預測等品質較差的深度圖。每個面元附屬於一個關鍵影格，只有在位姿圖上與參考關鍵影格相距少於 G_delta 條邊的面元才參與融合，使每一影格的融合時間與場景規模無關；位姿圖最佳化後，依各關鍵影格的位姿變化移動其面元，使整張地圖非剛性變形並保持全域一致。
本頁內容
CPU surfel fusion driven by a sparse SLAM pose graph: superpixel surfels (SLIC extended to intensity and depth with Huber-robust depth) are attached to keyframes, only surfels of pose-graph-local keyframes are fused so per-frame cost stays constant, and loop closures deform the map by moving each keyframe's surfels with its optimized pose; works with RGB-D, stereo or monocular predicted depth.
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 (synthetic ICL-NUIM input with ORB-SLAM2 in RGB-D mode)、Stereo camera (KITTI odometry; depth from PSMNet stereo matching, ORB-SLAM2 stereo mode)、Monocular camera with learned depth (KITTI left images with monocular depth prediction; handheld camera with MVDepthNet depth and VINS-Mono tracking) |
|---|---|
| 原文測試平台 | vehicle (KITTI odometry 00 and 05)、handheld (monocular camera used to build the map for quadrotor flights)、UAV (quadrotor flew aggressive autonomous paths planned on the reconstructed map)、simulation (ICL-NUIM synthetic room) |
| 狀態估計 | No pose estimation in the mapper: an external sparse visual SLAM (ORB-SLAM2 or VINS-Mono) provides each frame's pose, the reference keyframe and the optimized keyframe pose graph |
| 資料關聯 | Local surfels are projected into the current frame and matched to the surfel initialized from the superpixel at that pixel when depths agree within a disparity-derived bound and normals agree (dot product > 0.8); only surfels attached to keyframes within G_delta pose-graph edges of the reference keyframe (breadth-first search) are fused |
| 時間表示 | discrete keyframe poses |
| 去畸變 | 不適用 (camera input) |
| 迴圈閉合 | Provided by the localization system (ORB-SLAM2 or VINS-Mono); after loop closure previously built surfels are reused through new pose-graph edges |
| 全域最佳化 | Pose graph optimized by the external SLAM; surfels are deformed by applying each keyframe's pose change to its attached surfels (rigid per keyframe, non-rigid overall) |
| 地圖表示 | superpixel-based surfels (position, normal, intensity, weight, radius, update count, attached keyframe index) stored in a map database organized by keyframe |
| 先驗資訊 | none |
| 可輸出幾何 | globally consistent surfel map (shown as point clouds and meshes) |
| 計算需求 | CPU only; workstation with Intel i7-7700; average fusion time about 80 ms per frame on KITTI 00 (above 10 Hz), of which surfel fusion takes less than 6 ms regardless of scale; superpixel grid spacing 8 pixels |
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 | handheld monocular camera (model not reported) | 方法輸入 | 未標示 | depth from MVDepthNet, poses from VINS-Mono; map used for quadrotor flights | (Wang et al., 2019, Sec. VI-D) |
| 雙目相機 | KITTI stereo cameras (model not reported in the paper) | 資料集感測器 | KITTI odometry | depth maps from PSMNet stereo matching; left images also used for monocular depth prediction | (Wang et al., 2019, Sec. VI-B, VI-C) |
| 載具平台 | quadrotor (model not reported) | 方法輸入 | 未標示 | flew aggressive autonomous paths planned on the map reconstructed from the handheld monocular camera sequence (Sec. VI-D) | (Wang et al., 2019, Sec. VI-D) |
| 運算硬體 | workstation with an Intel i7-7700 | 執行運算平台 | 未標示 | all mapping on CPU | (Wang et al., 2019, Sec. VI) |
作者報告的優勢與限制
優勢
- Mean reconstruction error 0.7 to 1.1 cm on ICL-NUIM kt0 to kt3 with CPU only, similar to FlashFusion (Table I)
- Loop closure reduces kt3 error from 1.7 cm to 0.8 cm (Table I; Sec. VI-A)
- Real-time urban-scale reconstruction of KITTI 00 on CPU: about 80 ms per frame, surfel fusion under 6 ms regardless of scale (Sec. VI-B, Fig. 6)
- Memory grows with environment size rather than runtime because surfels are reused on revisits (Fig. 7)
- Map deformation keeps revisited fine obstacles consistent, unlike CHISEL given the same images and poses (Sec. VI-D, Fig. 8)
限制
- Depends on an external sparse SLAM system for poses, loop detection and pose-graph optimization (Sec. III, IV-B)
- Assumes keyframes within G_delta edges are locally consistent and moves surfels rigidly with their keyframe (Sec. IV-B, IV-C)
- Large-scale and flight experiments are evaluated by runtime, memory and qualitative figures only, without geometric ground truth (Sec. VI-B to VI-D)
- Most computation goes to superpixel extraction and surfel initialization rather than fusion (Sec. VI-B)
營建工程相關證據
原論文未在施工現場測試。corpus 中 Chen et al., 2025a 在 Unitree Go1 四足機器人上，以 Intel RealSense D455 的 RGB-D 資料執行本方法，對香港科技大學既有建築約 500 平方公尺的室內區域建圖，並以 Leica BLK360 地面雷射掃描為參考，經 ICP 對齊後的單向平均最近鄰距離為 0.1193 m、標準差 0.1356 m（該文 Table 8、Sec. 3.3.2）。該數值評估的是整體定位加建圖流程，並非單獨評估本方法。
原文驗證環境：公開基準、模擬
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 3 個比較組，合計 14 筆紀錄。
Wang et al., 2019 · Table I 本方法 8 筆
指標reconstruction accuracy (cm)
表格設定（擷取紀錄原文）：ICL-NUIM living room with simulated noise; reconstruction accuracy = mean difference between reconstructed model and ground-truth model; ORB-SLAM2 RGB-D mode for tracking, G_delta = 20; only FlashFusion and Ours run without GPU; 'Ours w/o loop' disables ORB-SLAM2 loop closure; comparator values coincide with FlashFusion Table III (Wang et al., 2019, Table I)
reconstruction accuracy (cm)，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。
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 Wang et al., 2019 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(Wang et al., 2019, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| BundleFusion | 0.5 cm | (Wang et al., 2019, Table I) |
| ElasticFusion | 0.7 cm | (Wang et al., 2019, Table I) |
| InfiniTAM [13] (Kaehler et al. ECCV 2016) | 1.3 cm | (Wang et al., 2019, Table I) |
| FlashFusion | 0.8 cm | (Wang et al., 2019, Table I) |
| Ours本方法原文提出 | 0.7 cm | (Wang et al., 2019, Table I) |
| Ours w/o loop (ablation: ORB-SLAM2 loop closure disabled)本方法 | 0.7 cm | (Wang et al., 2019, Table I) |
Yunus et al., 2021 · Table III 本方法 4 筆
指標reconstruction error (cm)
表格設定（擷取紀錄原文）：ICL-NUIM living room; reconstruction error of the point cloud generated from the surfels (cm); ElasticFusion and InfiniTAM need a GPU, DSM and ManhattanSLAM run on CPU; comparator provenance not stated (Yunus et al., 2021, Table III)
reconstruction error (cm)，ICL-NUIM · lr-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。
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 Yunus et al., 2021 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(Yunus et al., 2021, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| E-Fus [15] (ElasticFusion, IJRR version) | 0.7 cm | (Yunus et al., 2021, Table III) |
| InfiniTAM [41] (InfiniTAM v3 report) | 1.3 cm | (Yunus et al., 2021, Table III) |
| DSM [14] (Dense Surfel Mapping)本方法 | 0.7 cm | (Yunus et al., 2021, Table III) |
| Ours (ManhattanSLAM)原文提出 | 0.5 cm | (Yunus et al., 2021, Table III) |
Wang et al., 2019 · Text Sec.VI-B 本方法 2 筆
資料集與序列KITTI odometry · 00
表格設定（擷取紀錄原文）：KITTI odometry 00 reconstructed from PSMNet stereo depth with ORB-SLAM2 stereo tracking; per-frame fusion time including superpixel extraction and surfel initialization (Wang et al., 2019, Text Sec.VI-B)
average fusion time per frame，KITTI odometry · 00
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Wang et al., 2019 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Ours本方法原文提出硬體：Intel i7-7700 workstation (CPU only) | 80 ms有附註註記（擷取紀錄）：other: approximate value ('around 80 ms' per frame, Sec. VI-B) | (Wang et al., 2019, Sec. VI-B; Fig. 6) |
來源
Wang et al., 2019
（2019）Real-time Scalable Dense Surfel Mapping2019 International Conference on Robotics and Automation (ICRA), pp. 6919-6925
DOI 10.1109/icra.2019.8794101arXiv 1909.04250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：Real-time Scalable Dense Surfel Mapping (arXiv v1, posted after ICRA 2019) https://arxiv.org/abs/1909.04250
- 程式碼釋出：HKUST-Aerial-Robotics/DenseSurfelMapping https://github.com/HKUST-Aerial-Robotics/DenseSurfelMapping
程式碼：https://github.com/HKUST-Aerial-Robotics/DenseSurfelMapping（授權：not_stated (no LICENSE file; package.xml licence field reads 'TODO')）。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。