LDSO
LDSO 把直接稀疏里程計 DSO 擴充為具迴圈閉合的單目視覺 SLAM。它保留 DSO 以梯度選點的直接法追蹤，但讓部分選點偏向可重複的角點，只在關鍵影格上計算 ORB 描述子並建立詞袋資料庫以偵測迴圈。迴圈候選以 ORB 匹配與 RANSAC PnP 初始化，再同時最小化三維點對齊與二維重投影誤差求 Sim(3) 相對位姿；這些約束與滑動視窗的共視相對位姿一起放入 Sim(3) 位姿圖，以 g2o 最佳化修正旋轉、平移與尺度漂移，不做全域光束法平差。
本頁內容
LDSO extends direct sparse odometry (DSO) to monocular SLAM by biasing point selection toward repeatable corners for bag-of-words loop detection, estimating Sim(3) loop constraints from combined 3D and 2D geometric errors, and correcting drift with a Sim(3) pose graph fused with DSO's co-visibility constraints.
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 | ["TUM-Mono indoor and outdoor sequences (carrying mode not stated in the paper)", "EuRoC MAV sequences", "KITTI Odometry training sequences (platform not described in the paper)"] |
| 狀態估計 | DSO sliding-window photometric bundle adjustment (5 to 7 active keyframes, inverse-depth points, affine brightness and exposure) as the odometry front end; back end Sim(3) pose graph built from co-visibility relative poses of the window plus loop constraints, optimised with g2o while the current window poses are kept fixed (Secs. III-A, III-D) |
| 資料關聯 | direct photometric alignment for tracking; point selection keeps DSO's gradient-based pixels but favours Shi-Tomasi corners, for which ORB descriptors are computed on keyframes only and stored in a DBoW3 bag-of-words database (Secs. III-B, III-C) |
| 時間表示 | discrete keyframes |
| 去畸變 | 不適用 |
| 迴圈閉合 | DBoW3 query among marginalised keyframes, ORB matching and RANSAC PnP initial guess, then Gauss-Newton Sim(3) estimate minimising 3D point alignment and 2D reprojection terms using depths from the sliding window (Sec. III-C) |
| 全域最佳化 | Sim(3) pose-graph optimisation (g2o) over keyframes; no global bundle adjustment (Secs. III-D, V) |
| 地圖表示 | keyframe pose graph with sparse inverse-depth points from DSO; a point cloud map is shown before and after loop closure (Fig. 7) |
| 先驗資訊 | photometric camera calibration where the dataset provides it (TUM-Mono) |
| 可輸出幾何 | loop-corrected keyframe trajectory up to monocular scale and a sparse point cloud from DSO's points (Figs. 1, 7) |
| 計算需求 | laptop with Intel i7-4770HQ and 16 GB RAM (Ubuntu 18.04); LDSO point selection 0.0218 s per keyframe versus 0.0126 s for DSO; pose-graph optimisation runs in a separate thread (Sec. IV-D; Table II) |
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 | TUM-Mono camera (model not reported) | 資料集感測器 | TUM-Mono | 50 indoor and outdoor sequences with photometric calibration and start equal to end point | (Gao et al., 2018, Sec. IV-A) |
| 雙目相機 | EuRoC MAV stereo camera (model not reported; used monocularly) | 資料集感測器 | EuRoC MAV | stereo images with synchronised IMU and ground-truth trajectories; 11 sequences | (Gao et al., 2018, Sec. IV-B) |
| 運算硬體 | Intel i7-4770HQ laptop, 16GB RAM | 執行運算平台 | 未標示 | Ubuntu 18.04 | (Gao et al., 2018, Sec. IV-D) |
作者報告的優勢與限制
優勢
- Corner-biased point selection keeps DSO's odometry accuracy and robustness on TUM-Mono (Figs. 4-5; Sec. IV-A)
- On KITTI sequences with loops LDSO sharply reduces DSO's ATE, for example 126.7 m to 9.322 m on seq. 00 and 49.85 m to 5.1 m on seq. 05 (Table I)
- Accuracy comparable to ORB-SLAM2 on KITTI without global bundle adjustment, and it runs on seq. 01 where ORB-SLAM2 failed (Table I)
- More robust than ORB-SLAM2 on EuRoC, though ORB-SLAM2 is more accurate there (Sec. IV-B; Figs. 8-9, plotted)
- Descriptors are computed only for keyframes, keeping extra cost small (Sec. IV-D; Table II)
限制
- On KITTI seq. 08 the ATE stays large at 129.02 m, above DSO's 120.17 m
- the paper names only seq. 00, 05 and 07 as sequences with loops (Sec. IV-C
- Table I)
營建工程相關證據
論文未在營建工地測試；評估使用 TUM-Mono（室內外序列）、EuRoC MAV 與 KITTI Odometry 公開資料。直接法可利用白牆或弱紋理區的梯度，且迴圈閉合能修正單目尺度漂移，對低成本相機記錄施工進度有參考價值；但輸出僅為單目尺度的稀疏點雲，需外部尺度或感測器融合才能用於量測（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
Gao et al., 2018
（2018）LDSO: Direct Sparse Odometry with Loop Closure2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 2198-2204
DOI 10.1109/iros.2018.8593376arXiv 1808.01111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：LDSO: Direct Sparse Odometry with Loop Closure (arXiv v1) https://arxiv.org/abs/1808.01111
- 程式碼釋出：tum-vision/LDSO (project page vision.in.tum.de/research/vslam/ldso) https://github.com/tum-vision/LDSO
- 前身方法：Direct Sparse Odometry (DSO), the odometry front end that LDSO extends 10.1109/TPAMI.2017.2658577
程式碼：https://github.com/tum-vision/LDSO（授權：GPL-3.0 (LICENSE.txt)）。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。