DeepVO
DeepVO 是早期的端到端單眼視覺里程計：把相鄰兩張 RGB 影像疊合後送入以 FlowNet 預訓練權重初始化的卷積網路擷取運動特徵，再以兩層 LSTM 建模時間序列，直接迴歸每一時刻的六自由度位姿。方法不需特徵擷取、匹配、光束法平差，甚至不需相機校正，絕對尺度由訓練資料隱式學得。論文只在 KITTI 上驗證，平移漂移優於單眼 LIBVISO2，但仍明顯不如立體 LIBVISO2，且不產生地圖。
本頁內容
End-to-end monocular visual odometry that regresses 6-DoF poses directly from stacked consecutive RGB frames with a FlowNet-initialized CNN and a two-layer LSTM, learning absolute scale from KITTI training data; outputs poses only.
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) |
| 狀態估計 | end-to-end regression: a 9-layer CNN initialized from pretrained FlowNet extracts features from two stacked consecutive RGB frames, and two stacked LSTM layers (1000 hidden units each) output a 6-DoF pose per time step; trained with MSE on positions and Euler angles (orientation weight 100) (Sec. III) |
| 資料關聯 | none explicit; motion is learned implicitly from stacked image pairs without feature matching (Sec. III) |
| 時間表示 | discrete poses (one per frame) |
| 去畸變 | 不適用 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | none (poses only) |
| 先驗資訊 | supervised training on KITTI sequences with ground-truth poses; CNN initialized from a pretrained FlowNet model; absolute scale is learned from the training data (Sec. IV-A) |
| 可輸出幾何 | 6-DoF camera trajectory only; no map or point cloud |
| 計算需求 | implemented in Theano and trained on an NVIDIA Tesla K40 GPU; inference runtime not reported (Sec. 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 |
|---|---|---|---|---|---|
| 運算硬體 | NVIDIA Tesla K40 | 執行運算平台 | 未標示 | GPU used for training in Theano; inference runtime not reported | (Wang et al., 2017, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Mean translational drift 5.96% on KITTI test sequences versus 17.48% for monocular VISO2 (Table II)
- Recovers absolute scale without camera-height priors or post alignment to ground truth (Sec. IV-B)
- Needs no hand-designed VO modules or camera calibration (Sec. I, Sec. V)
限制
- Less accurate than stereo VISO2 (mean translational drift 1.89%) (Table II)
- Translational error grows at high speeds because training data above 50 km/h are scarce; sequence 12 shows large errors (Sec. IV-B)
- Prone to overfitting, especially in orientation, which degrades generalization (Sec. IV-A)
- Authors stress it is a complement, not a replacement, for geometry-based VO (Sec. V)
營建工程相關證據
論文未涉及營建場域；只以 KITTI 車載資料訓練與測試，輸出僅為軌跡而無地圖。其絕對尺度由訓練資料學得，換到不同相機或工地場景時是否仍成立並未驗證（推論），因此主要作為學習式里程計的歷史節點，而非可直接產生施工點雲的方法。
原文驗證環境：公開基準
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
Wang et al., 2017
（2017）DeepVO: Towards end-to-end visual odometry with deep Recurrent Convolutional Neural Networks2017 IEEE International Conference on Robotics and Automation (ICRA), pp. 2043-2050
DOI 10.1109/icra.2017.7989236arXiv 1709.08429
同儕審查已出版已讀全文近十年
相關版本
- postprint：arXiv:1709.08429v1 (ICRA 2017 paper posted after the conference) https://arxiv.org/abs/1709.08429