DPV-SLAM
DPV-SLAM 在稀疏影像區塊（patch）視覺里程計 DPVO 上加入兩種迴圈閉合，讓深度學習式單眼 SLAM 可在單張 GPU 上以穩定的影格速率運作。近距迴圈閉合依相機位置偵測重訪，只保留舊影格的區塊特徵並建立指向近期影格的單向邊，再以自製的 CUDA 區塊稀疏光束法平差做全域最佳化；DPV-SLAM++ 另以 DBoW2 影像檢索、特徵匹配與 RANSAC 加 Umeyama 估計 Sim(3) 漂移，於 CPU 執行位姿圖最佳化以修正尺度漂移。輸出僅為相機軌跡與稀疏點。
本頁內容
Extends the DPVO sparse-patch learned visual odometry to monocular SLAM on one GPU with proximity-based loop edges solved by block-sparse global BA and, in DPV-SLAM++, image-retrieval loop closure with Sim(3) pose-graph optimization; output is a trajectory and 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 | vehicle (KITTI)、UAV (EuRoC MAV dataset)、handheld (TUM RGB-D freiburg1)、simulation (TartanAir test set) |
| 狀態估計 | DPVO recurrent update operator predicts sparse patch-flow residuals and confidences; poses and patch inverse depths solved by bundle adjustment on the patch graph; a CUDA block-sparse BA performs global optimization with loop factors; DPV-SLAM++ adds a CPU Sim(3) pose-graph optimization solved by Levenberg-Marquardt (Sec. 3) |
| 資料關聯 | sparse, randomly selected p x p patches tracked by learned optical flow from correlation features; loop candidates by camera proximity (DPV-SLAM) and additionally by DBoW2 ORB image retrieval with off-the-shelf keypoint matching, structure-only BA and RANSAC plus Umeyama Sim(3) alignment (DPV-SLAM++) (Sec. 3.1-3.3) |
| 時間表示 | discrete poses (keyframes) |
| 去畸變 | 不適用 |
| 迴圈閉合 | proximity loop closure: uni-directional long-range edges from stored patches of old frames to recent frames, followed by global BA; optional classical loop closure (DPV-SLAM++) with image retrieval requiring consecutive detections and Sim(3) drift estimation (Sec. 3.2-3.3) |
| 全域最佳化 | global bundle adjustment over the patch graph mixed with odometry factors, limited to 1000 keyframes because cost grows quadratically; plus Sim(3) pose-graph optimization on the CPU for DPV-SLAM++ (Sec. 3.2-3.3, Sec. 5) |
| 地圖表示 | patch graph: sparse image patches with inverse depth attached to frames; only sparse 3D reconstruction (Sec. 3.1, Sec. 5) |
| 先驗資訊 | network trained only on synthetic data (Sec. 2); DPV-SLAM++ uses pretrained off-the-shelf keypoint detectors and matchers during loop closure (Sec. 3.3) |
| 可輸出幾何 | camera trajectory and sparse 3D points of tracked patches (Sec. 5) |
| 計算需求 | single GPU (RTX 3090 for timing); 50 FPS and 5.0 GB (DPV-SLAM) or 7.0 GB (DPV-SLAM++) on EuRoC, 39 FPS on KITTI, 27 FPS on TartanAir; classical loop closure adds about 2 GB GPU memory (Tables 1-3, Sec. 4, Sec. 5) |
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 used for all timing experiments | (Lipson et al., 2024, Sec. 4) |
作者報告的優勢與限制
優勢
- EuRoC average ATE 0.023 m versus 0.022 m for DROID-SLAM while running at 50 versus 20 FPS with 5.0 versus 20 GB GPU memory (ECCV version Table 3; arXiv v1 printed 0.024 m)
- Proximity loop closure reduces DPVO's EuRoC error from 0.105 to 0.023 m (about 4.5x, ECCV version Sec. 4) with a small speed and memory cost (Table 3)
- DPV-SLAM++ reaches 25.76 m average ATE on KITTI 00-10 where DROID-SLAM fails on two sequences (Table 2(b))
- Runs with a single GPU at a relatively consistent frame rate; no catastrophic failures reported across the four datasets (Sec. 1, Sec. 6)
限制
- Requires a GPU and provides only a sparse 3D reconstruction (Sec. 5)
- Global BA cost grows quadratically with pose count, so it is limited to 1000 keyframes (Sec. 5, Appendix A)
- Scale drift in outdoor environments; proximity-only loop detection fails under strong scale drift, as the KITTI results show (Sec. 4, Sec. 5)
- Image retrieval can give false positives, and the classical loop closure adds about 2 GB GPU memory (Sec. 5)
營建工程相關證據
論文未涉及營建場域；驗證資料為 KITTI、EuRoC、TUM RGB-D 與 TartanAir。它只輸出軌跡與稀疏點，無法直接產生可量測的稠密點雲；KITTI 上即使 DPV-SLAM++ 平均 ATE 仍約 25.76 m，顯示單眼尺度漂移在大範圍戶外仍是主要問題。可作為影像式重建或其他稠密建圖模組的位姿前端參考（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
Lipson et al., 2024
（2024）Deep Patch Visual SLAMComputer Vision - ECCV 2024 (Lecture Notes in Computer Science), pp. 424-440
DOI 10.1007/978-3-031-72627-9_24arXiv 2408.01654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:2408.01654v1 https://arxiv.org/abs/2408.01654
- 程式碼釋出：princeton-vl/DPVO (includes DPV-SLAM) https://github.com/princeton-vl/DPVO
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。