KinectFusion
KinectFusion 將 Kinect 深度串流即時融合到單一全域截斷符號距離函數(Truncated Signed Distance Function, TSDF)體素模型中,並以光線投射(raycasting)產生的模型表面預測,用由粗到細的 ICP(point-to-plane、投影式資料關聯)追蹤感測器位姿。追蹤對象是累積模型而非前一影格,因而在房間尺度內漂移有限。所有步驟皆可在 GPU 上平行化。
本頁內容
KinectFusion fuses depth frames into a global TSDF volume on the GPU and tracks each frame by coarse-to-fine point-to-plane ICP against a raycast prediction of the fused model.
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 | hand-held Kinect (Fig. 1; Sec. 1; experiment 5 with 560 x 4 free-moving frames)、Kinect fixed in place observing a tabletop scene on a turntable rotated through a full turn in about 19 s (560 frames) |
| 狀態估計 | coarse-to-fine point-to-plane ICP against the raycast model prediction over the bottom 3 pyramid levels with at most 4, 5 and 10 iterations (coarse to fine); small-angle linearisation gives per-correspondence 6x6 systems summed on the GPU by tree reduction and solved by Cholesky on the CPU; a null-space check and an increment-magnitude check switch the system into relocalisation mode |
| 資料關聯 | projective data association on bilateral-filtered depth (vertex and normal map pyramid, L = 3), rejecting pairs by vertex distance and normal-angle thresholds; raw depth, not the filtered depth, is fused into the TSDF |
| 時間表示 | discrete poses (every frame) |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | none explicit; loops are closed only implicitly by frame-to-model tracking (turntable loop-closing frames nearly overlap after one pass and more tightly after four passes); on tracking failure an interactive relocalisation asks the user to align the live depth frame with the prediction from the last known pose |
| 全域最佳化 | none |
| 地圖表示 | single fixed-extent dense TSDF volume in GPU memory storing truncated distance and weight (16 bits per component); projective TSDF with nearest-neighbour depth lookup and weighted running average, optional weight cap for moving-average reconstruction of dynamic scenes; 256^3 voxels in the turntable experiments, 64^3 to 512^3 evaluated |
| 先驗資訊 | none |
| 可輸出幾何 | dense TSDF surface rendered by raycasting; mesh export not reported in sections read |
| 計算需求 | commodity GPU (model not reported); TSDF update above 65 gigavoxels per second (about 2 ms per full 512^3 volume); tracking at the 30 Hz Kinect frame rate; constant-time tracking and mapping for a given voxel resolution; 64^3 volume with every 6th frame shows graceful degradation |
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 | Kinect | 方法輸入 | 未標示 | structured-light depth sensor; on-board ASIC produces an 11-bit 640x480 depth map at 30 Hz; conservative range about 0.4 to 8 m used for raycasting; only depth used | (Newcombe et al., 2011b, Sec. 2.1; Sec. 3.4) |
| 運算硬體 | commodity GPU (model not reported) | 執行運算平台 | 未標示 | all tracking and mapping on GPU; TSDF update above 65 gigavoxels per second | (Newcombe et al., 2011b, Abstract; Sec. 3.3) |
| 其他 | turntable | 參考或真值量測 | 未標示 | tabletop scene rotated through a full rotation in about 19 s (560 frames) with the Kinect fixed, equivalent to a precise circular sensor path | (Newcombe et al., 2011b, Sec. 4.1) |
作者報告的優勢與限制
優勢
- Constant-time tracking and mapping in room-sized scenes with limited drift (abstract)
- Uses only depth, so robust to indoor lighting conditions (Sec. 4.3)
限制
- Main failure case: a large planar scene filling most of the view leaves three DOF unconstrained, causing drift or failure (Sec. 4.3)
- Works well for medium-sized rooms with volumes of <=7 m3 (as printed); reconstructing the interior of a whole building would raise memory problems and drift appearing as misalignment upon loop closures (Sec. 6 Conclusions)
- Relocalisation is interactive; efficient automatic relocalisation in large models is an open challenge (Sec. 3.5; Sec. 6)
- Kinect depth has holes where materials or scene structures do not reflect IR light, on very thin structures and at glancing incidence angles, and suffers motion blur under fast motion (Sec. 2.1)
營建工程相關證據
論文未於營建現場測試。作者指出大面積平面會造成 ICP 三自由度不受約束,並將整棟建築重建列為尚未解決的記憶體與漂移問題;這兩點直接對應室內牆面、樓板與大空間掃描情境(推論連結)。另外 Kinect 深度在不反射紅外光的材料、細長結構與掠射角表面會產生空洞(Sec. 2.1),工地的鋼筋、玻璃與斜視牆面可能受影響(推論);論文的迴圈實驗為轉盤定性展示,未提供數值精度。
原文驗證環境:受控實驗
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 5 個比較組,合計 17 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 1 組列在最後,並連到性能比較頁。
Kerl et al., 2013 · Table III 本方法 10 筆
指標RMSE of absolute trajectory error
表格設定(擷取紀錄原文):RMSE of absolute trajectory error (m) on TUM RGB-D sequences versus RGB-D SLAM (Engelhard, Endres et al.), MRSMap and the PCL KinectFusion implementation (KinFu); dashes = not available (Kerl et al., 2013, Table III)
RMSE of absolute trajectory error,TUM RGB-D · 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。
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 Kerl et al., 2013 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(Kerl et al., 2013, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Ours (DVO-SLAM)原文提出 | 0.011 m | (Kerl et al., 2013, Table III) |
| RGB-D SLAM [2], [31] | 0.014 m | (Kerl et al., 2013, Table III) |
| MRSMap [11] | 0.013 m | (Kerl et al., 2013, Table III) |
| KinFu (PCL KinectFusion) [5]本方法 | 0.026 m | (Kerl et al., 2013, Table III) |
Kähler et al., 2015 · Table 1 本方法 4 筆
指標average computation time per frame
表格設定(擷取紀錄原文):Average computation time per frame over the teddy sequence (Kinect for XBOX 360, 640x480 colour and disparity, no IMU) for three visualisation strategies of InfiniTAM and for the KinectFusion [14] and voxel hashing [16] implementations (Kähler et al., 2015, Table 1)
average computation time per frame,authors' teddy sequence · teddy
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 Kähler 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(Kähler et al., 2015, Table 1)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| KinectFusion implementation [14](Nvidia Titan X)本方法 | 26.15 ms | (Kähler et al., 2015, Table 1(a)) |
| KinectFusion implementation [14](Intel Core i7-5960X)本方法 | 502.69 ms | (Kähler et al., 2015, Table 1(a)) |
| Voxel hashing implementation [16]硬體:Nvidia Titan X | 25.87 ms | (Kähler et al., 2015, Table 1(a)) |
Ghadimzadeh Alamdari et al., 2025 · Table 2 本方法 1 筆
指標Result (run outcome)
資料集與序列Luleå SubT tunnel dataset (Koval et al. 2022)
表格設定(擷取紀錄原文):Run outcome ('Result' column) of each reviewed vision-based method on the Luleå tunnel test dataset; '+' marks methods not integrated with ROS; the '*' (incompatible with VLP-16) symbol is printed on almost every row (Ghadimzadeh Alamdari et al., 2025, Table 2)
Result (run outcome),Luleå SubT tunnel dataset (Koval et al. 2022)
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。
- 失敗
- 未執行
- 未報告(沒有數值,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 Ghadimzadeh Alamdari et al., 2025 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(Ghadimzadeh Alamdari et al., 2025, Table 2)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Mono-SLAM | 無數值失敗註記(擷取紀錄):failed (feature detection and tracking) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| PTAM | 無數值失敗註記(擷取紀錄):failed (initialization for ground floor) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| S-PTAM | 無數值未執行註記(擷取紀錄):未執行 (authors could not run the code) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| OV2SLAM | 無數值未執行註記(擷取紀錄):未執行 (authors could not run the code) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| ORB-SLAM (footnote 1) | 無數值未報告註記(擷取紀錄):success (footnote 1: authors could not run ORB-SLAM 3, so the original ORB-SLAM was used) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| DTAM | 無數值未執行註記(擷取紀錄):未執行 (no publicly available repository) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| LSD-SLAM | 無數值失敗註記(擷取紀錄):failed (feature tracking) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| SVO | 無數值失敗註記(擷取紀錄):failed (tracking) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| DSO | 無數值未執行註記(擷取紀錄):未執行 (authors could not run the code) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| Kinetic Fusion本方法 | 無數值未執行註記(擷取紀錄):未執行 (no publicly available repository) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| Dense visual SLAM | 無數值失敗註記(擷取紀錄):failed (feature detection and tracking) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| Elastic Fusion SLAM | 無數值未執行註記(擷取紀錄):未執行 (inconsistent repository) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| Realtime onboard VI estimation | 無數值未執行註記(擷取紀錄):未執行 (no publicly available repository) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| Multi-sensor fusion | 無數值未執行註記(擷取紀錄):未執行 (no publicly available repository) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| SOFT-SLAM | 無數值未執行註記(擷取紀錄):未執行 (no publicly available repository) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| MSCKF | 無數值失敗註記(擷取紀錄):failed (tracking) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| ROVIO | 無數值未執行註記(擷取紀錄):未執行 (authors could not run the code) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| OKVIS | 無數值失敗註記(擷取紀錄):failed (tracking) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| VIORB | 無數值未執行註記(擷取紀錄):未執行 (inconsistent repository) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| S-MSCKF | 無數值未執行註記(擷取紀錄):未執行 (no publicly available repository) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| VINS-Mono | 無數值未報告註記(擷取紀錄):success | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| STCM-SLAM | 無數值未執行註記(擷取紀錄):未執行 (no publicly available repository) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| Kimera | 無數值失敗註記(擷取紀錄):failed (feature detection and tracking) | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
| Yolo-SLAM | 無數值未報告註記(擷取紀錄):other: Result cell reads 'SLAM for dynamic environments'; no run outcome stated | (Ghadimzadeh Alamdari et al., 2025, Table 2) |
Newcombe et al., 2011b · Text Sec. 1 本方法 1 筆
指標tracking frame rate
資料集與序列live Kinect input
表格設定(擷取紀錄原文):Tracking rate stated in text; tracking and mapping run at the Kinect frame rate in constant time for a given voxel resolution (Newcombe et al., 2011b, Text Sec. 1)
tracking frame rate,live Kinect input
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 Newcombe et al., 2011b 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 |
|---|---|---|
| KinectFusion本方法原文提出硬體:commodity GPU (model not reported) | 30 Hz | (Newcombe et al., 2011b, Sec. 1; Sec. 4.2) |
其他比較組
列出其餘 1 個比較組
來源
Newcombe et al., 2011b
(2011)KinectFusion: Real-time dense surface mapping and tracking2011 10th IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pp. 127-136
DOI 10.1109/ismar.2011.6092378
同儕審查已出版已讀全文經典查證後修正