PTAM
PTAM 將相機追蹤(tracking)與建圖(mapping)拆成兩個平行執行緒:追蹤執行緒以地圖點重投影估計每張影像的位姿,建圖執行緒則對關鍵影格(keyframe)執行計算量較大的光束法平差(bundle adjustment, BA)。此設計讓即時系統可以使用原本多用於離線 SfM 的批次最佳化。作者將其定位為小型擴增實境工作區,並未支援大範圍探索。
本頁內容
PTAM separates frame-rate tracking from keyframe-based bundle-adjusted mapping in parallel threads, enabling batch optimisation in a real-time monocular system for small AR workspaces.
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 | monocular camera |
|---|---|
| 原文測試平台 | handheld |
| 狀態估計 | tracking thread: pose update by ten iterations of reweighted least squares on a Tukey-biweight reprojection objective; mapping thread: Levenberg-Marquardt bundle adjustment with a Tukey M-estimator, run globally or locally over the newest keyframe and its four nearest keyframes (Sec. 5.4, 6.3) |
| 資料關聯 | FAST-10 corners on a four-level image pyramid; affine-warped 8x8 patches searched by zero-mean SSD at FAST corners around reprojected map points, 50 coarse-level points then up to 1000 points per frame; new map points triangulated by epipolar search against the nearest keyframe (Sec. 5.1, 5.3, 5.5, 6.2) |
| 時間表示 | discrete poses (keyframes) |
| 去畸變 | 不適用 |
| 迴圈閉合 | no dedicated loop-detection module; authors state the system is not designed to close large loops, although the BA mapping can absorb loops when the camera is placed near the map boundary (Sec. 8) |
| 全域最佳化 | bundle adjustment over keyframes and map points in the mapping thread |
| 地圖表示 | sparse point-feature map with keyframes |
| 先驗資訊 | no prior map; user-assisted five-point stereo initialisation in which metric scale is set by assuming a 10 cm camera translation between the first two keyframes and the dominant plane found by RANSAC is placed at z = 0 (Sec. 6.1) |
| 可輸出幾何 | keyframe poses and sparse 3D point features |
| 計算需求 | Intel Core 2 Duo 2.66 GHz desktop PC running Linux, C++ with libCVD and TooN (Sec. 6.5); about 19.2 ms to track one frame with a 4000-point map (Table 1); mean global bundle adjustment 380 ms, 1.7 s and 6.9 s for maps of 2-49, 50-99 and 100-149 keyframes versus 170, 270 and 440 ms for local bundle adjustment (Table 2) |
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 | Unibrain Fire-i | 方法輸入 | 未標示 | video camera with 2.1 mm wide-angle lens; 640x480 YUV411 frames at 30 Hz, converted to 8-bit greyscale | (Klein & Murray, 2007, Sec. 5.1) |
| 運算硬體 | Intel Core 2 Duo 2.66 GHz desktop PC | 執行運算平台 | 未標示 | dual-core processor, Linux, C++ with libCVD and TooN | (Klein & Murray, 2007, Sec. 6.5) |
作者報告的優勢與限制
優勢
- add: on a synthetic 600-frame sequence (camera path 18.2 m) the trajectory standard deviation from ground truth after 7-DOF alignment was 6 mm versus 135 mm for an EKF-SLAM implementation, with near-constant 20 ms tracking (Sec. 7.3)
限制
- add: the map is only a point cloud
- after extracting the dominant plane the system makes little effort to extract further geometric understanding, and patch normals are initialised parallel to the image plane and not optimised (Sec. 6.5, 8.2)
營建工程相關證據
論文未報告營建或建築量測測試;設計目標為桌面或房間角落等小型 AR 工作區。對本文僅作為關鍵影格 BA 架構的技術源頭。
原文驗證環境:受控實驗、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 8 個比較組,合計 39 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 4 組列在最後,並連到性能比較頁。
Mur-Artal et al., 2015 · Table III 本方法 16 筆
指標Absolute KeyFrame Trajectory RMSE
表格設定(擷取紀錄原文):TUM RGB-D keyframe ATE RMSE (cm), median over 5 executions; ORB-SLAM, PTAM (two manually chosen initial frames) and LSD-SLAM (first 10 keyframes discarded) aligned with Sim(3); RGBD-SLAM trajectories from the benchmark website aligned with SE(3), Sim(3) value in brackets; X = tracking lost (Mur-Artal et al., 2015, Table III)
Absolute KeyFrame Trajectory RMSE,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 Mur-Artal 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(Mur-Artal et al., 2015, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| ORB-SLAM原文提出 | 0.9 cm | (Mur-Artal et al., 2015, Table III) |
| PTAM本方法 | 1.15 cm | (Mur-Artal et al., 2015, Table III) |
| LSD-SLAM | 9 cm | (Mur-Artal et al., 2015, Table III) |
Klein & Murray, 2007 · Table 2 本方法 6 筆
表格設定(擷取紀錄原文):Mean bundle adjustment time by map size (keyframes); timings vary with map size and scene structure (Klein & Murray, 2007, Table 2)
local bundle adjustment time,typical timings from live operation (no specific sequence named; Sec. 7.2) · map with 2-49 keyframes
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 Klein & Murray, 2007 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 |
|---|---|---|
| PTAM (proposed system)本方法原文提出硬體:Intel Core 2 Duo 2.66 GHz desktop PC, Linux (Sec. 6.5) | 170 ms | (Klein & Murray, 2007, Table 2) |
Klein & Murray, 2007 · Table 1 本方法 5 筆
資料集與序列own live video (Sec. 7.1) · typical frame, map of M=4000 points (Table 1 is not tied to a named sequence)
表格設定(擷取紀錄原文):Tracking time for a typical frame broken down by step, map size M=4000 (Sec. 7.1, Table 1) (Klein & Murray, 2007, Table 1)
tracking timing: Keyframe preparation,own live video (Sec. 7.1) · typical frame, map of M=4000 points (Table 1 is not tied to a named sequence)
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 Klein & Murray, 2007 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 |
|---|---|---|
| PTAM (proposed system)本方法原文提出硬體:Intel Core 2 Duo 2.66 GHz desktop PC, Linux (Sec. 6.5) | 2.2 ms | (Klein & Murray, 2007, Table 1) |
Engel et al., 2014 · Fig. 9 本方法 4 筆
指標absolute trajectory RMSE (cm)
表格設定(擷取紀錄原文):Result table printed as Fig. 9: absolute trajectory RMSE (cm) on TUM RGB-D and two simulated sequences from Handa et al.; LSD-SLAM also lists keyframes created; x = tracking failure, '-' = no data; [14] and [7] use sensor depth; alignment not stated. The ORB-SLAM authors report being unable to reproduce the PTAM fr2/xyz value (orbslam2015, Sec. VIII-B) (Engel et al., 2014, Fig. 9)
absolute trajectory RMSE (cm),TUM RGB-D · fr2/desk
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
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 Engel et al., 2014 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(Engel et al., 2014, Fig. 9)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| semi-dense mono-VO [9] | 13.5 cm | (Engel et al., 2014, Fig. 9) |
| keypoint-based mono-SLAM [15] (PTAM)本方法 | 無數值失敗註記(擷取紀錄):failed (tracking failure) | (Engel et al., 2014, Fig. 9) |
其他比較組
來源
Klein & Murray, 2007
(2007)Parallel Tracking and Mapping for Small AR Workspaces2007 6th IEEE and ACM International Symposium on Mixed and Augmented Reality (ISMAR), pp. 1-10 as registered in Crossref (citing works often give pp. 225-234; not verified on IEEE Xplore)
DOI 10.1109/ismar.2007.4538852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:PTAM-GPL https://github.com/Oxford-PTAM/PTAM-GPL
程式碼:https://github.com/Oxford-PTAM/PTAM-GPL(授權:GPLv3 (LICENSE file of Oxford-PTAM/PTAM-GPL; the PTAM project page links this 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。