C-blox
C-blox 把場景表示為一組相互重疊的 TSDF 子體積（subvolume），每個子體積固定附著在 ORB-SLAM2 的一個關鍵影格上。迴圈閉合後，只要以最佳化後的關鍵影格位姿更新子體積座標系，就能修正稠密地圖，不必重新整合深度影像。為避免子體積數量隨軌跡長度無限增加，作者先以地標共視圖找出可能重複觀測同一區域的子體積對，再從光束法平差資訊矩陣計算兩者相對定位的條件共變異，只有定位品質足夠時才把兩個子體積融合，藉此限制地圖成長。整個系統只用 CPU，並在 voxblox 中加入多執行緒快速整合器，可在無人機上的 Intel NUC 即時執行。
本頁內容
CPU TSDF submapping: overlapping voxblox TSDF subvolumes are anchored to ORB-SLAM2 keyframes so loop closures correct the dense map by moving subvolume frames, and map growth is limited by fusing covisible subvolume pairs whose relative-pose conditional covariance, recovered from the bundle-adjustment information matrix, is small.
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 | Stereo camera pair of a VI-sensor (global shutter, tightly synchronized) for ORB-SLAM2 tracking on the MAV、RGB-D camera Intel RealSense D415 (coloured pointclouds) for dense integration on the MAV、Synthetic RGB-D input (ICL-NUIM) and simulated RGB plus noiseless depth (CARLA) in the evaluations |
|---|---|
| 原文測試平台 | UAV (hexacopter on a DJI F550 frame, two flights in an indoor industrial area)、simulation (CARLA car drives through two synthetic cities; ICL-NUIM synthetic living room) |
| 狀態估計 | Modified ORB-SLAM2 (keyframe bundle adjustment of feature re-projection errors with Huber cost) supplies camera poses; each subvolume is rigidly attached to a keyframe and moved when keyframe poses are re-optimized; no dense tracking against the TSDF |
| 資料關聯 | Sparse ORB features for tracking; subvolume-fusion candidates from a landmark covisibility graph (edge weight = number of shared landmarks), accepted only if the relative localization quality q = 1/||Sigma_i|j|| from the bundle-adjustment information matrix exceeds a threshold (Schur complement, constrained AMD reordering and Cholesky-based covariance recovery) |
| 時間表示 | discrete keyframe poses |
| 去畸變 | 不適用 (depth camera input) |
| 迴圈閉合 | ORB-SLAM2 loop detection and bundle adjustment; the dense map is corrected by updating subvolume base frames with the optimized keyframe poses |
| 全域最佳化 | Sparse keyframe bundle adjustment in ORB-SLAM2; dense map corrected rigidly per subvolume, without re-integration of depth frames |
| 地圖表示 | collection of overlapping TSDF subvolumes (voxblox) with no spatial partitioning; a new subvolume starts after a maximum number of keyframes or a large sparse-map change; redundant well-localized subvolumes are fused by trilinear interpolation |
| 先驗資訊 | none |
| 可輸出幾何 | TSDF subvolume collection and fused mesh; voxel size 0.02 m on ICL-NUIM and 0.5 m on CARLA |
| 計算需求 | CPU only; on-board Intel NUC Core i7-7567U with tracking and integration at 10 Hz; the multi-threaded fast integrator cut depth integration from 93 ms to 25 ms per frame (Sec. V-C); desktop Intel Core i7-7820X for ICL-NUIM |
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 |
|---|---|---|---|---|---|
| 雙目相機 | Visual-Inertial (VI-)sensor [28] | 方法輸入 | 未標示 | tightly synchronized stereo images from a pair of global shutter cameras; used for camera tracking at 10 Hz | (Millane et al., 2018, Sec. V-C) |
| RGB-D 相機 | Intel RealSense D415 Depth Camera歸入：Intel RealSense D415 | 方法輸入 | 未標示 | coloured pointclouds integrated at 10 Hz; extrinsics to the tracked camera calibrated offline with Kalibr | (Millane et al., 2018, Sec. V-C) |
| 載具平台 | hexacopter based on a DJI F550 frame | 方法輸入 | 未標示 | px4 autopilot for low-level attitude stabilization; all high-level computation on board | (Millane et al., 2018, Sec. V-C) |
| 運算硬體 | Intel NUC Core i7-7567U | 執行運算平台 | 未標示 | on board; runs the proposed approach entirely in real time | (Millane et al., 2018, Sec. V-C) |
| 運算硬體 | desktop PC with Intel Core i7-7820X CPU at 3.60 GHz and Nvidia Titan Xp GPU | 執行運算平台 | 未標示 | used for the ICL-NUIM reconstruction comparison; the GPU was used only by ElasticFusion | (Millane et al., 2018, Sec. V-A) |
作者報告的優勢與限制
優勢
- Median surface RMSE 0.015 m on ICL-NUIM with 0.02 m voxels, close to voxblox with ground-truth poses (0.013 m) without using a GPU (Table I)
- On two CARLA drives the median RMSE was 0.68 m (fusion off) and 0.72 m (fusion on) versus 2.09 m for voxblox with uncorrected ORB-SLAM2 poses (Table II)
- Subvolume fusion reduced the map to 55% (l0) and 68% (l1) of the allocated blocks without fusion (Table II; Sec. V-B)
- Runs in real time entirely on the MAV computer (Sec. V-C)
限制
- Slightly higher error than GPU-based ElasticFusion on ICL-NUIM kt0 to kt2 (Table I; Sec. V-A)
- Compression depends on the path and on the fusion threshold q; earlier fusion gives smaller maps but can cost accuracy (Sec. V-B)
- The CARLA 'ground truth' is itself a voxblox reconstruction with 0.25 m voxels from ground-truth poses (Sec. V-B)
- Requires 3D observations (stereo or depth) so that ORB-SLAM2 has metric scale (Sec. IV-B)
- The fast integrator terminates rays early and enforces a time budget, trading completeness of updates for speed (Sec. IV-C)
營建工程相關證據
論文未在施工現場測試。實機實驗為無人機在室內工業區飛行兩次（可見管線結構），只有定性結果，沒有幾何參考量測（Fig. 1、Sec. V-C）。corpus 中 Vizzo et al., 2021 在 Table I 以 C-blox 作為 TSDF 基準。子體積附著於稀疏 SLAM 關鍵影格、迴圈閉合後整塊移動的作法，可作為樓層尺度室內掃描事後修正的參考（推論）。
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 16 筆紀錄。
Millane et al., 2018 · Table II 本方法 10 筆
表格設定（擷取紀錄原文）：CARLA simulated drives l0 and l1 through two synthetic cities; reference geometry is a voxblox reconstruction with ground-truth poses and 0.25 m voxels; evaluated systems use 0.5 m voxels; tracking and integration at 10 Hz; 'Voxblox (ORB-SLAM Poses)' has no dense-map correction after loop closure (Millane et al., 2018, Table II)
RMSE (m)，CARLA (simulated) · l0
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 Millane 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。
資料來源作者報告值(Millane et al., 2018, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Voxblox (GT Poses) | 0.52 m | (Millane et al., 2018, Table II) |
| Voxblox (ORB-SLAM Poses) | 2.12 m | (Millane et al., 2018, Table II) |
| Ours (subvolume fusion OFF, ablation)本方法 | 0.59 m | (Millane et al., 2018, Table II) |
| Ours (subvolume fusion ON)本方法原文提出 | 0.66 m | (Millane et al., 2018, Table II) |
Millane et al., 2018 · Table I 本方法 5 筆
指標RMSE (m)
表格設定（擷取紀錄原文）：ICL-NUIM living room (synthetic, noisy depth); RMSE between mesh vertices (or surfel centres) and the closest ground-truth surface point after alignment (method not stated); voxel size 0.02 m for voxblox-based systems; kt3 is the only sequence with a loop closure; the printed ElasticFusion median (0.001 m) is inconsistent with its per-sequence values (Millane et al., 2018, Table I)
RMSE (m)，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 Millane 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。
資料來源作者報告值(Millane et al., 2018, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ElasticFusion | 0.006 m | (Millane et al., 2018, Table I) |
| Voxblox (GT Poses) | 0.01 m | (Millane et al., 2018, Table I) |
| Ours (C-blox)本方法原文提出 | 0.011 m | (Millane et al., 2018, Table I) |
Millane et al., 2018 · Text Sec.V-C 本方法 1 筆
指標average time required to integrate depth data
資料集與序列authors' MAV industrial flights · f0 and f1
表格設定（擷取紀錄原文）：Average time to integrate depth data per frame in the industrial MAV environment, original voxblox integrator versus the proposed fast integrator (Millane et al., 2018, Text Sec.V-C)
average time required to integrate depth data，authors' MAV industrial flights · f0 and f1
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 Millane 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。
資料來源作者報告值(Millane et al., 2018, Text Sec.V-C)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| voxblox original integrator硬體：Intel NUC Core i7-7567U (on board) | 93 ms | (Millane et al., 2018, Sec. V-C) |
| C-blox fast integrator本方法原文提出硬體：Intel NUC Core i7-7567U (on board) | 25 ms | (Millane et al., 2018, Sec. V-C) |
來源
Millane et al., 2018
（2018）C-blox: A Scalable and Consistent TSDF-based Dense Mapping Approach2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 995-1002
DOI 10.1109/iros.2018.8593427arXiv 1710.07242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：C-blox (arXiv v1 to v3) https://arxiv.org/abs/1710.07242
- 程式碼釋出：ethz-asl/cblox https://github.com/ethz-asl/cblox
- 後續方法：Voxgraph (RA-L 2020), which cites C-blox as the authors' previous work and reuses its indoor MAV dataset 10.1109/LRA.2019.2953859
程式碼：https://github.com/ethz-asl/cblox（授權：BSD-3-Clause (LICENSE file checked)）。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。