DM-VIO
DM-VIO 是單目視覺慣性里程計，以 DSO 的直接光度光束法平差為核心，加入 IMU 預積分並把尺度與重力方向作為顯式變數持續最佳化。作者提出延遲邊際化：另外維護一個延遲 100 個關鍵影格才邊際化的因子圖，可在其中加入 IMU 因子做位姿圖光束法平差（PGBA），以完整的光度不確定度初始化 IMU，並重新推進該圖得到含 IMU 資訊的邊際化先驗；尺度大幅改變時也能替換邊際化先驗。另以動態光度權重在影像品質差時提高 IMU 比重。
本頁內容
DM-VIO is a direct monocular VIO that keeps a delayed marginalisation graph to run pose graph bundle adjustment for IMU initialisation with the full photometric uncertainty, to readvance an IMU-informed marginalisation prior, and to replace it when scale changes, while scale and gravity remain optimised in the main system.
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、IMU |
|---|---|
| 原文測試平台 | UAV (EuRoC MAV)、handheld (TUM-VI, large-scale indoor and outdoor)、vehicle (4Seasons) |
| 狀態估計 | direct (DSO-based) photometric visual-inertial bundle adjustment over up to 8 active keyframes with IMU preintegration, dynamic photometric weight, and explicit scale and gravity-direction variables; Schur-complement partial marginalisation with FEJ plus a second delayed marginalisation graph (delay 100) used for pose graph bundle adjustment (PGBA) IMU initialisation and marginalisation replacement; Levenberg-Marquardt with SIMD photometric code and GTSAM (Sec. III) |
| 資料關聯 | direct: photometric residuals of sparse DSO points over a neighbourhood pattern with affine brightness and exposure; no feature matching (Sec. III-B) |
| 時間表示 | discrete keyframes with IMU preintegration between keyframes (Sec. III-B) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none (odometry; loop closure and map reuse named as future work) |
| 全域最佳化 | none |
| 地圖表示 | sparse point cloud of inverse-depth points hosted in active keyframes (Fig. 1 point clouds) |
| 先驗資訊 | priors on first pose and gravity direction; IMU noise parameters from calibration or data sheets (Secs. III-B, IV-C) |
| 可輸出幾何 | metric-scale camera and IMU trajectory with sparse point cloud (Fig. 1) |
| 計算需求 | real-time mode on a MacBook Pro 2013 (i7 at 2.3 GHz) without GPU: tracking 10.34 ms per frame and keyframe processing 53.67 ms on average; delayed marginalisation overhead 0.44 ms (0.8%) in the keyframe thread (Sec. IV; IV-A) |
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 | TUM-VI handheld visual-inertial device (model not reported in this paper) | 資料集感測器 | TUM-VI | large-scale indoor and outdoor handheld sequences, including sliding down a tube | (von Stumberg & Cremers, 2022, Sec. IV-B) |
| 相機 | 4Seasons visual-inertial sensor (model not reported) | 資料集感測器 | 4Seasons | well time-synchronised; bottom 96 pixels cropped because of the car hood; IMU noise read from the data-sheet Allan variance plot | (von Stumberg & Cremers, 2022, Sec. IV-C) |
| 相機 | EuRoC MAV camera and IMU (models not reported in this paper) | 資料集感測器 | EuRoC MAV | monocular images and IMU data recorded by a flying drone | (von Stumberg & Cremers, 2022, Secs. I, IV-A) |
| 運算硬體 | MacBook Pro 2013 (i7 at 2.3GHz) | 執行運算平台 | 未標示 | real-time mode without GPU; same machine as in the VI-DSO paper | (von Stumberg & Cremers, 2022, Sec. IV) |
| 運算硬體 | Intel Core i7-7700K at 4.2GHz desktop | 執行運算平台 | 未標示 | used only to run ORB-SLAM3 (not supported on macOS) | (von Stumberg & Cremers, 2022, Sec. IV) |
作者報告的優勢與限制
優勢
- Lowest average EuRoC RMSE among the compared VIO systems (0.069 m, versus 0.072 m for stereo Basalt and 0.089 m for VI-DSO) with average scale error 0.6% (Table I)
- On TUM-VI best on 16 sequences with mean normalised drift 0.472% versus 0.939% for Basalt (Table II)
- Outperforms stereo-inertial ORB-SLAM3 and Basalt on 4Seasons despite monocular input and no loop closure (Sec. IV-C; Fig. 5, plotted)
- Delayed marginalisation costs only about 0.44 ms (0.8%) in the keyframe thread (Sec. IV-A)
限制
- Odometry only: ORB-SLAM3 with loop closure is more accurate on some TUM-VI sequences (Sec. IV-B; Fig. 4)
- Scale is not observable during constant-velocity motion, so monocular VIO initialisation remains difficult in automotive scenes (Secs. I, IV-C)
- Large drift remains on long TUM-VI outdoor sequences (for example 123.24 m on outdoors1, 2656 m long) (Table II)
- 4Seasons required cropping the car hood and modifying the visual initialiser (zero prior on x-y translation, keyframe threshold) (Sec. IV-C)
營建工程相關證據
論文未在營建工地測試；評估涵蓋 EuRoC（無人機）、TUM-VI（大範圍室內外手持資料，含沿管道滑下的序列）與 4Seasons（車輛）。單目相機加 IMU 即可得到公制尺度且漂移低的軌跡，對以手持或頭戴裝置在施工中建物內巡檢定位有參考價值；但只輸出稀疏點雲且無迴圈閉合，長距離戶外仍有數公尺至數十公尺漂移（推論）。
原文驗證環境：公開基準、獨立參考量測
報告的性能數據
性能數據仍在分批查證，目前尚未收錄此方法的報告值。
來源
von Stumberg & Cremers, 2022
（2022）DM-VIO: Delayed Marginalization Visual-Inertial OdometryIEEE Robotics and Automation Letters, 7(2):1408-1415
DOI 10.1109/lra.2021.3140129arXiv 2201.04114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：DM-VIO: Delayed Marginalization Visual-Inertial Odometry (arXiv v1, accepted RA-L version) https://arxiv.org/abs/2201.04114
- 程式碼釋出：lukasvst/dm-vio (project page vision.in.tum.de/dm-vio; supplementary with ablations and runtime) https://github.com/lukasvst/dm-vio
- 前身方法：VI-DSO: Direct sparse visual-inertial odometry using dynamic marginalization (ICRA 2018), cited as [6] and compared in Table I 10.1109/ICRA.2018.8462905
程式碼：https://github.com/lukasvst/dm-vio（授權：GPL-3.0 (LICENSE file)）。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。