CHISEL
CHISEL 在 Google Tango 手機與平板上,只用行動裝置的 CPU 即時建立房屋尺度(300 平方公尺以上)的 TSDF 稠密重建,不使用 GPU 通用運算。作者採用 Nießner 等人的空間雜湊兩層結構,把 16×16×16 體素的區塊(chunk)依視錐裁剪結果動態配置,未被更新的區塊即回收,只處理含有表面的空間以節省記憶體與運算。針對 Tango 深度感測器雜訊大、更新率只有 3 至 6 Hz 的問題,加入依雜訊模型調整的動態截斷距離與空間雕刻(space carving)。定位以裝置內建的視覺慣性里程計為輸入,再以掃描對 TSDF 模型的 ICP 修正短距漂移;網格則以增量 marching cubes 依區塊延遲產生,系統本身沒有迴圈閉合。
本頁內容
CPU-only house-scale TSDF reconstruction on Google Tango devices: spatially hashed 16^3-voxel chunks allocated by frustum culling and garbage-collected, dynamic truncation and space carving from a trained depth-noise model, visual-inertial odometry corrected by scan-to-TSDF ICP, and lazy per-chunk incremental meshing; 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 | Google Tango 'Peanut' phone: projective depth sensor (6 Hz), 120 deg wide-angle tracking camera (60 Hz), 4 MP colour camera (30 Hz), six-axis gyroscope and accelerometer、Google Tango 'Yellowstone' tablet: projective depth sensor (3 Hz), same tracking camera, 4 MP colour camera (30 Hz)、Kinect RGB-D data of the Freiburg (TUM) benchmark for memory experiments |
|---|---|
| 原文測試平台 | handheld (Tango phone and tablet; office building floor, 175 m corridor, apartment, outdoor night scene) |
| 狀態估計 | Onboard visual-inertial odometry (EKF fusing wide-angle camera and inertial data with 2D feature tracking at 60 Hz; the paper refers to Kottas et al. and the MSCKF for details) with sparse keypoint mapping used as a black box, incrementally corrected by scan-to-model ICP against the TSDF (residual = TSDF value at each transformed point, gradient by central differences); corrective transforms accumulated per scan |
| 資料關聯 | Implicit point-to-TSDF association via the signed distance value and its gradient (first-order projection onto the zero level set) |
| 時間表示 | discrete poses |
| 去畸變 | 不適用 (depth camera) |
| 迴圈閉合 | none online; Fig. 7 shows a corridor corrected only after offline bundle adjustment |
| 全域最佳化 | none online (offline bundle adjustment used only for the Fig. 7 corridor illustration) |
| 地圖表示 | dynamic spatially hashed TSDF: chunks of 16 x 16 x 16 voxels in a hash map, allocated on frustum intersection and garbage-collected when not updated; each voxel stores a 16-bit fixed-point SDF and 16-bit weight plus 8-bit RGB and colour weight; dynamic truncation from a trained depth-noise model; space carving |
| 先驗資訊 | none |
| 可輸出幾何 | per-chunk triangle meshes from incremental marching cubes (lazy, asynchronous), coloured by trilinear interpolation; map savable to disk; 2 to 3 cm voxels on the devices |
| 計算需求 | CPU only on the mobile device (no general-purpose GPU computing); single-scan fusion 62 to 200 ms on the Tango tablet and 14 to 58 ms on a desktop depending on the fusion mode (Fig. 9e) |
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 |
|---|---|---|---|---|---|
| 行動掃描設備 | Google Tango 'Yellowstone' tablet | 方法輸入 | 未標示 | 4 GB RAM, quad-core CPU, Nvidia Tegra K1 graphics, 120 deg FOV tracking camera at 60 Hz, projective depth sensor at 3 Hz, 4 megapixel colour sensor at 30 Hz | (Klingensmith et al., 2015, Sec. IV-A) |
| 行動掃描設備 | Google Tango 'Peanut' mobile phone | 方法輸入 | 未標示 | 2 GB RAM, quad-core CPU, six-axis gyroscope and accelerometer, 120 deg wide-angle tracking camera at 60 Hz, projective depth sensor at 6 Hz, 4 megapixel colour sensor at 30 Hz | (Klingensmith et al., 2015, Sec. IV-A) |
| RGB-D 相機 | Kinect | 資料集感測器 | Freiburg (TUM) RGB-D benchmark | sensor loops around a central desk in a roughly 12 by 12 m room; the paper writes only 'Kinect', manufacturer inferred from the product name | (Klingensmith et al., 2015, Sec. IV-D) |
| 運算硬體 | desktop machine (model not reported) | 執行運算平台 | 未標示 | used only for the fusion-time comparison in Fig. 9e | (Klingensmith et al., 2015, Sec. IV-C; Fig. 9e) |
| 其他 | motion capture system (model not reported) | 參考或真值量測 | Freiburg (TUM) RGB-D benchmark | ground-truth poses of the Freiburg dataset | (Klingensmith et al., 2015, Sec. IV-D) |
作者報告的優勢與限制
優勢
- Real-time house-scale reconstruction at 2 to 3 cm voxel resolution entirely on a mobile CPU (Abstract; Sec. IV-B)
- Spatial hashing needed about a tenth of the fixed-grid memory as the explored space grew, never more than 47 MB for a 15 m hallway (Sec. IV-D, Fig. 5a)
- Per-frame meshing and update of 102 ms and 128 ms with spatial hashing versus 2067 ms and 3769 ms for a 256^3 fixed grid (Table II)
- Space carving strongly reduces noise artifacts around object silhouettes and removes briefly seen moving objects (Sec. III-E, IV-C, Fig. 9)
限制
- No global consistency; the map drifts over time, about 5 m at the end of a roughly 175 m corridor with VIO and dense alignment only (Sec. III-J, Sec. V, Fig. 7)
- Resolution of 2 to 3 cm is much coarser than GPU TSDF systems that reach sub-centimetre voxels (Sec. V)
- Tango depth arrives at only 3 to 6 Hz, too slow for depth-only tracking (Sec. III-J)
- Projection mapping at 3 cm causes aliasing on surfaces nearly parallel to the viewing axis; the constant weighting approximation degrades surfaces in noisy areas (Sec. III-B, IV-C)
- Quantitative results cover timing and memory only; no surface accuracy against ground truth is reported (Sec. IV)
營建工程相關證據
論文未在施工現場測試。實例包括整層辦公大樓、約 175 公尺走廊、公寓與夜間戶外場景,皆以手持 Tango 裝置即時掃描(Fig. 1、7、8)。在行動裝置上即時建立 2 至 3 公分解析度的樓層模型,符合低成本手持室內掃描的需求;但系統會累積漂移(走廊末端約 5 公尺),且論文沒有以幾何真值評估精度,用於量測用途前需另以控制點校核(推論)。
原文驗證環境:公開基準、已完工建築
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 50 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。
Klingensmith et al., 2015 · Fig. 9e 本方法 32 筆
資料集與序列authors' Room (apartment) dataset · Room
表格設定(擷取紀錄原文):Time to fuse a single depth scan on the 'Room' (apartment, Fig. 1b) dataset for each fusion mode; mean and standard deviation printed as 'mean +- std' (Klingensmith et al., 2015, Fig. 9e)
single scan fusion time (ms.),authors' Room (apartment) dataset · Room
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 Klingensmith et al., 2015 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(Klingensmith et al., 2015, Fig. 9e)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| CHISEL: Raycast, no colour, no carving(desktop (model not reported))本方法原文提出 | 14 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Raycast, no colour, no carving(Google Tango 'Yellowstone' tablet)本方法原文提出 | 62 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Raycast, colour, no carving(desktop (model not reported))本方法原文提出 | 20 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Raycast, colour, no carving(Google Tango 'Yellowstone' tablet)本方法原文提出 | 80 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Raycast, no colour, carving(desktop (model not reported))本方法原文提出 | 53 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Raycast, no colour, carving(Google Tango 'Yellowstone' tablet)本方法原文提出 | 184 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Raycast, colour, carving(desktop (model not reported))本方法原文提出 | 58 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Raycast, colour, carving(Google Tango 'Yellowstone' tablet)本方法原文提出 | 200 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Projection mapping, no colour, no carving(desktop (model not reported))本方法原文提出 | 33 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Projection mapping, no colour, no carving(Google Tango 'Yellowstone' tablet)本方法原文提出 | 106 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Projection mapping, colour, no carving(desktop (model not reported))本方法原文提出 | 39 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Projection mapping, colour, no carving(Google Tango 'Yellowstone' tablet)本方法原文提出 | 125 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Projection mapping, no colour, carving(desktop (model not reported))本方法原文提出 | 34 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Projection mapping, no colour, carving(Google Tango 'Yellowstone' tablet)本方法原文提出 | 116 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Projection mapping, colour, carving(desktop (model not reported))本方法原文提出 | 40 ms | (Klingensmith et al., 2015, Fig. 9e) |
| CHISEL: Projection mapping, colour, carving(Google Tango 'Yellowstone' tablet)本方法原文提出 | 128 ms | (Klingensmith et al., 2015, Fig. 9e) |
Han & Fang, 2018 · Table IV 本方法 6 筆
表格設定(擷取紀錄原文):Efficiency comparison between CPU-based CHISEL and FlashFusion on TUM fr3/office at three voxel resolutions; time per operation in ms (Han & Fang, 2018, Table IV)
TSDF Fusion (ms),TUM RGB-D · fr3/office (5mm voxels)
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 Han & Fang, 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(Han & Fang, 2018, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| CHISEL本方法硬體:Intel Core i7 7700 @3.6 GHz (CPU only) | 483 ms | (Han & Fang, 2018, Table IV) |
| FlashFusion原文提出硬體:Intel Core i7 7700 @3.6 GHz (CPU only) | 3.6 ms | (Han & Fang, 2018, Table IV) |
Klingensmith et al., 2015 · Table I 本方法 4 筆
資料集與序列Freiburg (TUM) RGB-D benchmark · Freiburg 5m (desk loop)
表格設定(擷取紀錄原文):Voxel statistics for the Freiburg 5 m depth-frustum reconstruction; share of the bounding box per voxel class (culled voxels are not stored) (Klingensmith et al., 2015, Table I)
% of Bounding Box (Unknown Culled),Freiburg (TUM) RGB-D benchmark · Freiburg 5m (desk loop)
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 Klingensmith et al., 2015 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 |
|---|---|---|
| CHISEL spatially hashed TSDF本方法原文提出 | 77% | (Klingensmith et al., 2015, Table I) |
Klingensmith et al., 2015 · Table II 本方法 4 筆
資料集與序列authors' Room (apartment) dataset · Room
表格設定(擷取紀錄原文):Per-frame mesh generation and TSDF update (colorization, space carving, projection mapping) on the 'Room' dataset; platform not stated (update time equals the tablet value in Fig. 9e) (Klingensmith et al., 2015, Table II)
Meshing Time (ms.),authors' Room (apartment) dataset · Room
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 Klingensmith et al., 2015 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(Klingensmith et al., 2015, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| 256^3 Fixed Grid (baseline)硬體:not stated | 2067 ms | (Klingensmith et al., 2015, Table II) |
| 16^3 Spatial Hashing (CHISEL)本方法原文提出硬體:not stated | 102 ms | (Klingensmith et al., 2015, Table II) |
其他比較組
來源
Klingensmith et al., 2015
(2015)Chisel: Real Time Large Scale 3D Reconstruction Onboard a Mobile Device using Spatially Hashed Signed Distance FieldsRobotics: Science and Systems XI (RSS 2015), 原文未報告 (paper 40)
DOI 10.15607/rss.2015.xi.040程式碼
同儕審查已出版已讀全文經典查證後修正
相關版本
- 程式碼釋出:personalrobotics/OpenChisel (open-source ROS reference implementation; the paper omits the link for double-blind review) https://github.com/personalrobotics/OpenChisel
程式碼:https://github.com/personalrobotics/OpenChisel(授權:MIT (stated in open_chisel/package.xml and source headers; no LICENSE file in the repository))。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。