DPVO
DPVO 是深度學習式單眼視覺里程計，把 DROID-SLAM 的稠密光流改為只追蹤稀疏影像區塊（patch）。每張影格隨機取樣區塊，循環更新網路依相關特徵、時間向卷積與訊息傳遞預測區塊軌跡修正量與信心權重，再由可微分光束法平差在滑動視窗內更新位姿與區塊逆深度。整個網路只以合成資料 TartanAir 訓練，在 RTX-3090 上平均每秒 60 影格、約 4.9 GB 記憶體，但不含迴圈閉合，輸出僅為軌跡與稀疏點。
本頁內容
Learned monocular visual odometry that tracks sparse random image patches with a recurrent update operator and differentiable bundle adjustment in a sliding window, matching or beating dense-flow DROID-VO at a fraction of the memory; no loop closure and only sparse points.
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 | UAV (EuRoC MAV dataset)、simulation (TartanAir, ICL-NUIM)、原文未報告 (TUM RGB-D platform not described; the paper notes erratic motion and blur) |
| 狀態估計 | recurrent update operator (correlation, 1D temporal convolution, softmax aggregation, transition block, factor head) predicts 2D patch-trajectory revisions and confidences; a differentiable bundle adjustment layer applies two Gauss-Newton iterations with the Schur complement to camera poses and patch inverse depths; poses of all but the last 10 keyframes are fixed (Sec. 3.1, Sec. 3.3, Appendix F) |
| 資料關聯 | sparse image patches at random locations (96 per frame by default, 48 in the fast setting) tracked by learned correlation features against frames within distance r in a bipartite patch graph (Sec. 3, Sec. 4) |
| 時間表示 | discrete poses (keyframes; relative poses stored for removed keyframes) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none (visual odometry only; DPV-SLAM later adds loop closure) |
| 全域最佳化 | none; sliding-window optimization over the last 10 keyframes (Sec. 3.3) |
| 地圖表示 | patch graph of sparse fronto-parallel patches with inverse depth; sparse 3D reconstruction (Sec. 3, Fig. 1, Appendix E) |
| 先驗資訊 | network trained entirely on synthetic TartanAir data (Sec. 3.2, Appendix D) |
| 可輸出幾何 | camera trajectory and sparse 3D points of tracked patches (Fig. 1, Fig. B) |
| 計算需求 | RTX-3090: default averages 60 FPS with 4.9 GB, fast setting 120 FPS with 2.5 GB, frame rate above 48 FPS for 95% of frames; training 3.5 days on one RTX-3090 (Sec. 1, Sec. 3.2, Figs. 8 and 10) |
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 |
|---|---|---|---|---|---|
| 運算硬體 | RTX-3090 | 執行運算平台 | 未標示 | GPU for runtime measurements and for training (single GPU, 3.5 days) | (Teed et al., 2023, Sec. 1, Sec. 3.2) |
作者報告的優勢與限制
優勢
- EuRoC average ATE 0.105 m versus 0.186 m for DROID-VO (Table 2)
- TartanAir test average ATE 0.21 versus 0.33 for DROID-SLAM with global optimization (Table 1)
- Averages 60 FPS with 4.9 GB versus 8.7 GB for DROID-VO, and the frame rate stays nearly constant regardless of motion (Sec. 1, Figs. 8 and 10)
- Learning-based and did not fail on any TUM RGB-D fr1 sequence, unlike ORB-SLAM3 and DSO (Table 3)
限制
- Odometry only: no loop closure or global correction, so drift accumulates (Sec. 2; DPV-SLAM Sec. 1)
- Only sparse 3D points are produced (Fig. 1, Appendix E) (inference for dense-mapping use)
- Treating all frames as keyframes is slower than necessary for slow or still camera motion (Appendix B)
- Initialization needs camera motion of at least 8 pixels average flow (Sec. 3.3)
- Reported by the DPV-SLAM authors: DPVO suffers the same performance issues as DROID-SLAM on outdoor data (DPV-SLAM Sec. 2)
營建工程相關證據
論文未涉及營建場域；評估資料為 TartanAir、EuRoC、TUM RGB-D 與 ICL-NUIM，均未含施工現場。DPVO 只輸出軌跡與稀疏點，且無迴圈閉合，無法單獨產生可量測的點雲；其價值在於作為低記憶體的學習式位姿前端，已被 DPV-SLAM 等後續系統沿用（推論）。
原文驗證環境：公開基準、模擬
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
Teed et al., 2023
（2023）Deep Patch Visual OdometryAdvances in Neural Information Processing Systems 36 (NeurIPS 2023), pp. 39033-39051
DOI 10.52202/075280-1696arXiv 2208.04726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：arXiv:2208.04726v2 https://arxiv.org/abs/2208.04726
- 程式碼釋出：princeton-vl/DPVO https://github.com/princeton-vl/DPVO
- extension：Deep Patch Visual SLAM (DPV-SLAM), ECCV 2024 https://doi.org/10.1007/978-3-031-72627-9_24
程式碼：https://github.com/princeton-vl/DPVO（授權：MIT (LICENSE file checked)）。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。