LO-Net
LO-Net 將相鄰兩幀 LiDAR 點雲以圓柱投影編碼成含距離與強度的資料矩陣，以孿生（Siamese）卷積網路直接迴歸 6 自由度相對位姿；網路內以距離加權的鄰點外積計算逐點法向量（非以可學習權重估計），並同時學習動態物遮罩（mask），以遮罩加權的法向量幾何一致性損失約束訓練。之後利用法向量挑選平滑區點、以遮罩排除移動物，再以點到平面（point-to-plane）的掃描對地圖（scan-to-map）配準精化位姿以降低累積漂移。此方法需要真值位姿做監督訓練，且只在 KITTI 與 Ford 車載資料上評估。
本頁內容
A supervised end-to-end network regresses relative LiDAR pose from cylindrical range-image encodings, using in-network analytic normals and a learned dynamic-object mask, and a point-to-plane scan-to-map module uses those outputs to refine odometry.
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 | Velodyne HDL-64 3D LiDAR (KITTI and Ford data; Sec. 4, 4.5) |
|---|---|
| 原文測試平台 | vehicle |
| 狀態估計 | supervised Siamese CNN regressing relative 6-DoF pose (translation + quaternion) from two scans, followed by iterative point-to-plane scan-to-map refinement (Sec. 3.3, 3.5) |
| 資料關聯 | implicit in the network (cylindrical range/intensity matrix input) with a mask-weighted normal-consistency loss; normals are computed inside the network by range-weighted cross products of four grid neighbours plus moving-average smoothing (Eq. 3), not by learned weights; mapping uses point-to-plane correspondences to map points, selecting smooth-area points by a convolution over the normal channels and excluding masked points (Sec. 3.2, 3.4, 3.5) |
| 時間表示 | discrete poses (scan-to-scan); linear motion interpolation used inside the mapping step (Sec. 3.5) |
| 去畸變 | mapping module removes motion distortion by linear interpolation of the LO-Net odometry before scan-to-map matching (Sec. 3.5) |
| 迴圈閉合 | none (not implemented for any method in the experiments, Sec. 4.2) |
| 全域最佳化 | none |
| 地圖表示 | sliding point map holding the last n_m = 100 transformed scans (Sec. 3.5, Sec. 4 implementation details) |
| 先驗資訊 | supervised training with ground-truth poses (KITTI 00-06); mask sub-network pre-trained on the KITTI 3D object detection dataset (Sec. 4) |
| 可輸出幾何 | 6-DoF trajectory and an accumulated point map shown in figures; map export is not described (Fig. 4) |
| 計算需求 | about 80.1 ms per scan on KITTI Seq. 00: 8.5 ms data preparation (CPU), 10.2 ms inference (GTX 1080 Ti GPU), 61.4 ms mapping (CPU, Intel Core i7 3.4 GHz) (Sec. 4.5, Table 4) |
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 |
|---|---|---|---|---|---|
| LiDAR | Velodyne HDL-64歸入：Velodyne HDL-64E | 資料集感測器 | KITTI odometry | 64 laser beams, 10 Hz, about 1.3 million points/second; encoded as 64 x 1800 matrices cropped to 1792 | (Li et al., 2019, Sec. 4 implementation details; Sec. 4.5) |
| LiDAR | Velodyne HDL-64歸入：Velodyne HDL-64E | 資料集感測器 | Ford Campus Vision and Lidar | horizontally scanning 3D lidar mounted on top of a vehicle; rotates at 10 Hz | (Li et al., 2019, Sec. 4.1; Sec. 4.5) |
| 慣性量測單元（IMU） | IMU (model not named) | 參考或真值量測 | Ford Campus Vision and Lidar | Ford ground-truth trajectories generated from IMU readings | (Li et al., 2019, Supplementary Fig. 18 caption; Sec. 4.1) |
| GNSS 接收器 | IMU/GPS (models not named) | 參考或真值量測 | KITTI odometry | ground-truth poses for KITTI sequences 00-10 | (Li et al., 2019, Sec. 4.1) |
| 運算硬體 | NVIDIA 1080 Ti | 執行運算平台 | 未標示 | GPU for training and for 10.2 ms inference | (Li et al., 2019, Sec. 4 implementation details; Sec. 4.5; Table 4) |
| 運算硬體 | Intel Core i7 3.4GHz 4-core CPU | 執行運算平台 | 未標示 | data preparation 8.5 ms and mapping 61.4 ms per scan | (Li et al., 2019, Sec. 4.5; Table 4) |
作者報告的優勢與限制
優勢
- model trained only on KITTI was applied to the Ford dataset without fine-tuning and reported the best average errors among the methods the authors ran (Sec. 4.2, Table 1)
- learned mask highlights moving objects and vegetation so they receive less weight (Sec. 4.4, Fig. 9)
- total processing under the 0.1 s scan period on a GPU plus CPU platform (Sec. 4.5, Table 4)
限制
- requires ground-truth poses for training, which the authors say limits application scenarios (Sec. 5)
- point clouds must be encoded into data matrices instead of processed directly (Sec. 5)
- LOAM baseline numbers were produced by the authors after modifying the original code; they differ from the LOAM paper's numbers (Sec. 4.2, Table 1 note)
- supervised learned LO methods rely on ground truth and target driving data with limited rotational motion (Nubert et al., 2021 Sec. II)
- (inference) no test on handheld, legged, indoor or construction data, so domain transfer to construction sites is unknown
營建工程相關證據
未在營建工地、完工建物、隧道或基礎設施測試；只使用 KITTI 與 Ford 車載資料。其監督式訓練需真值位姿，在工地取得真值的成本未被討論。補充材料說明以 fireConv 等輕量模組降低參數量，目的在於讓網路能在行動機器人或背包系統等低效能平台執行，但論文未在這類平台實測。
原文驗證環境：公開基準
報告的性能數據
以下是原文作者報告的性能數值（author-reported results），不是本研究重新量測的結果。每張圖只並列同一個比較組（comparison group，同一張表、同一組實驗設定）內的方法；不同比較組之間的數值不可直接比較，也不構成排名。
本方法共出現在 10 個比較組，合計 119 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 6 組列在最後，並連到性能比較頁。
Li et al., 2019 · Table 1 本方法 32 筆
表格設定（擷取紀錄原文）：KITTI odometry metric: t_rel = average translational RMSE (%) and r_rel = average rotational RMSE (deg/100 m) over 100-800 m lengths. LO-Net trained on KITTI 00-06 and tested on 07-10 and on Ford without fine-tuning; loop closure disabled for all methods. LOAM values outside brackets come from the authors' modified re-run; bracketed values are quoted from the LOAM paper [45]. Velas et al. values quoted from [35] (r_rel and Ford NA). ICP variants run with PCL. Truncated: per-sequence rows 00-06 (training sequences) omitted; the mean over them (mean-dagger) is kept. (Li et al., 2019, Table 1)
t_rel: average translational RMSE (%) on length of 100 m-800 m，KITTI odometry · 07 (not used for training)
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Li et al., 2019 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Li et al., 2019, Table 1)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ICP-po2po (PCL) | 5.17% | (Li et al., 2019, Table 1) |
| ICP-po2pl (PCL) | 1.55% | (Li et al., 2019, Table 1) |
| GICP [30] | 0.64% | (Li et al., 2019, Table 1) |
| CLS [34] | 1.04% | (Li et al., 2019, Table 1) |
| LOAM [45] (authors' modified re-run) | 0.69% | (Li et al., 2019, Table 1) |
| Velas et al. [35] (values from [35]) | 1.77% | (Li et al., 2019, Table 1) |
| LO-Net本方法原文提出 | 1.7% | (Li et al., 2019, Table 1) |
| LO-Net+Mapping本方法原文提出 | 0.56% | (Li et al., 2019, Table 1) |
Yokozuka et al., 2021 · Table III 本方法 22 筆
指標KITTI stats translation [%]
表格設定（擷取紀錄原文）：KITTI odometry sequences 00-10; KITTI stats translation error (%) averaged over 100-800 m segments with the benchmark code; rows without loop closure; LiTAMIN2 with 3 m voxels; LOAM (from [10]), LO-Net and DeepLO values copied from their papers (Yokozuka et al., 2021, Table III)
KITTI stats translation [%]，KITTI odometry · 00 (4541 frames)
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Yokozuka et al., 2021 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Yokozuka et al., 2021, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LiTAMIN2 (ICP+Cov), without loop closure原文提出 | 0.78% | (Yokozuka et al., 2021, Table III) |
| LiTAMIN2 (ICP), without loop closure原文提出 | 0.75% | (Yokozuka et al., 2021, Table III) |
| LiTAMIN [2], without loop closure | 0.91% | (Yokozuka et al., 2021, Table III) |
| SuMa (Frame-to-Frame) | 2.11% | (Yokozuka et al., 2021, Table III) |
| SuMa (Frame-to-Model), without loop closure | 0.72% | (Yokozuka et al., 2021, Table III) |
| LeGO-LOAM | 2.17% | (Yokozuka et al., 2021, Table III) |
| hdl graph slam | 3.92% | (Yokozuka et al., 2021, Table III) |
| LOAM (open source, run by authors) | 1.92% | (Yokozuka et al., 2021, Table III) |
| LOAM (from [10]) | 0.78% | (Yokozuka et al., 2021, Table III) |
| LO-Net (Frame-to-Frame)本方法 | 1.47% | (Yokozuka et al., 2021, Table III) |
| LO-Net (Frame-to-Model)本方法 | 0.78% | (Yokozuka et al., 2021, Table III) |
Pan et al., 2021 · Table II 本方法 16 筆
表格設定（擷取紀錄原文）：KITTI odometry ATE [%] and ARE [deg/100m], averaged over 100-800 m segments; baseline values from original papers and KITTI leaderboard; * = with loop closure; time in s per frame (Pan et al., 2021, Table II)
ATE [%] (average translation error)，KITTI odometry · 00-10 mean
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Pan et al., 2021 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Pan et al., 2021, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOAM [10] | 0.84% | (Pan et al., 2021, Table II) |
| IMLS-SLAM [11] | 0.52% | (Pan et al., 2021, Table II) |
| MC2SLAM [13] | 0.52% | (Pan et al., 2021, Table II) |
| S4-SLAM [26]* | 0.92% | (Pan et al., 2021, Table II) |
| PSF-LO [27] | 0.74% | (Pan et al., 2021, Table II) |
| SUMA++ [16]* | 0.7% | (Pan et al., 2021, Table II) |
| LiTAMIN2 [51]* | 0.85% | (Pan et al., 2021, Table II) |
| LO-Net [18]本方法 | 0.83% | (Pan et al., 2021, Table II) |
| FALO [25] | 1% | (Pan et al., 2021, Table II) |
| LoDoNet [28] | 1.27% | (Pan et al., 2021, Table II) |
| MULLS-LO(mc)原文提出 | 0.49% | (Pan et al., 2021, Table II) |
| MULLS-SLAM(mc)*原文提出 | 0.52% | (Pan et al., 2021, Table II) |
| MULLS-LO(s1)原文提出 | 2.57% | (Pan et al., 2021, Table II) |
| MULLS-SLAM(m1)*原文提出 | 0.77% | (Pan et al., 2021, Table II) |
| MULLS-SLAM(m5)*原文提出 | 0.6% | (Pan et al., 2021, Table II) |
| MULLS-SLAM(s5m5)*原文提出 | 0.61% | (Pan et al., 2021, Table II) |
Wang et al., 2021c · Table 1 本方法 13 筆
表格設定（擷取紀錄原文）：KITTI odometry, trained on 00-06 (marked *) and tested on 07-10; trel = average translational RMSE (%) over 100-800 m subsequences; rows other than LOAM w/o mapping and Ours are copied from LO-Net [10]; LOAM is a full system with mapping, others are odometry only (Wang et al., 2021c, Table 1)
trel (average translational RMSE, %)，KITTI odometry · 07 (test)
只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Wang et al., 2021c 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。
資料來源作者報告值(Wang et al., 2021c, Table 1)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Full LOAM [31] | 0.69% | (Wang et al., 2021c, Table 1) |
| ICP-po2po | 5.17% | (Wang et al., 2021c, Table 1) |
| ICP-po2pl | 1.55% | (Wang et al., 2021c, Table 1) |
| GICP [19] | 0.64% | (Wang et al., 2021c, Table 1) |
| CLS [21] | 1.04% | (Wang et al., 2021c, Table 1) |
| Velas et al. [22] | 1.77% | (Wang et al., 2021c, Table 1) |
| LO-Net [10]本方法 | 1.7% | (Wang et al., 2021c, Table 1) |
| DMLO [11] | 0.73% | (Wang et al., 2021c, Table 1) |
| LOAM w/o mapping (published code run by authors) | 10.87% | (Wang et al., 2021c, Table 1) |
| Ours (PWCLO-Net)原文提出 | 0.6% | (Wang et al., 2021c, Table 1) |
其他比較組
來源
Li et al., 2019
（2019）LO-Net: Deep Real-Time Lidar Odometry2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8465-8474 (Crossref); the CVF open-access PDF footer shows 8473-8482
DOI 10.1109/cvpr.2019.00867arXiv 1904.08242
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：LO-Net: Deep Real-time Lidar Odometry (arXiv v1 2019-04-17, v2 2020-01-17) https://arxiv.org/abs/1904.08242