PWCLO-Net
PWCLO-Net 是直接以原始 3D 點雲學習的監督式 LiDAR 里程計。它借用光流網路的金字塔、變形與代價體(PWC)結構:兩幀點雲先經共享權重的點特徵金字塔,再以注意力代價體建立軟對應;可訓練的嵌入遮罩為每個點加權,以回歸整體位姿並壓低動態物體與雜草等不可靠點的影響。估得的位姿用來變形前一幀,再由粗到細逐層修正位姿與遮罩。它只做逐幀里程計,沒有地圖或迴圈閉合。
本頁內容
Supervised deep LiDAR odometry on raw 3D points using a pyramid, warping and cost-volume (PWC) structure: an attentive point cost volume gives soft correspondences, a trainable embedding mask weights points to regress the pose and suppress outliers, and pose warp-refinement refines the estimate coarse to fine; frame-to-frame 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 | 3D LiDAR point coordinates only (KITTI Velodyne; reflectance not used) (Sec. 4.1) |
|---|---|
| 原文測試平台 | vehicle (KITTI) |
| 狀態估計 | supervised end-to-end network: siamese point feature pyramid (set conv with farthest point sampling and kNN), attentive point cost volume, trainable embedding mask, and three pose warp-refinement modules that refine a quaternion and translation over four levels; multi-level supervised loss with learnable weighting (Sec. 3; Sec. 4.2) |
| 資料關聯 | soft correspondences from an attentive cost volume between two consecutive frames (8192 randomly sampled points each) instead of explicit point matching; the embedding mask down-weights dynamic and irregular points (Sec. 3.2; Sec. 3.3; Sec. 4.2; Sec. 5.3) |
| 時間表示 | discrete frame-to-frame relative poses |
| 去畸變 | not described; KITTI clouds used as provided |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | none (frame-to-frame odometry without a map) |
| 先驗資訊 | requires ground-truth poses for supervised training (main setting trained on KITTI 00-06); inputs cropped to a 30 m x 30 m square and ground below 0.55 m removed (Sec. 4.1) |
| 可輸出幾何 | relative 6-DoF pose per frame pair; trajectory by chaining |
| 計算需求 | training and evaluation on a single NVIDIA RTX 2080Ti with TensorFlow 1.9.0; inference time is not reported in the main paper (Sec. 4.2) |
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 LiDAR (KITTI; model not specified in the paper) | 資料集感測器 | KITTI odometry | XYZ and reflectance provided; only XYZ used; transformed to the left camera frame | (Wang et al., 2021c, Sec. 4.1) |
| 運算硬體 | NVIDIA RTX 2080Ti歸入:NVIDIA RTX 2080 Ti | 執行運算平台 | 未標示 | single GPU, TensorFlow 1.9.0, used for training and evaluation | (Wang et al., 2021c, Sec. 4.2) |
作者報告的優勢與限制
優勢
- Mean errors on KITTI 07-10 (trained on 00-06) of 1.085% and 0.490 deg/100m, lower than LO-Net (1.748%, 0.793) and the full LOAM values from [10] (1.145%, 0.498); DMLO has a lower translational mean (1.008%) but a higher rotational mean (0.538) (Table 1)
- Better than LodoNet, DeepPCO and the unsupervised method of Cho et al. under each of their training splits (Tables 2-4)
- Embedding mask gives low weight to moving cars, cyclists, bushes and weeds without a separate mask network (Sec. 5.3; Fig. 8)
- Ablations show large gains from the cost volume and pose warp-refinement (Table 5)
限制
- Odometry only; combining the mask with mapping optimization is left to future work (Sec. 6)
- Trained and tested only on KITTI; generalization to other sensors or environments is not evaluated (inference from Sec. 4-5)
- Inputs are cropped to 30 m x 30 m and ground points below 0.55 m are removed (Sec. 4.1)
- Most baseline numbers are copied from LO-Net [10]; only LOAM without mapping was rerun (Table 1 caption)
- Inference runtime not reported in the main paper (Sec. 4.2)
營建工程相關證據
與施工沒有直接證據:只在 KITTI 道路資料上以監督方式訓練與測試,且輸入裁成車輛周圍 30 m 見方並移除地面。學習式里程計在工地這種訓練資料缺乏、場景快速變化的環境是否能泛化,論文沒有驗證。它在 NeRF-LOAM 的比較表中被列為基準(Deng et al., 2023)。
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 4 個比較組,合計 25 筆紀錄。
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] | 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)本方法原文提出 | 0.6% | (Wang et al., 2021c, Table 1) |
Wang et al., 2021c · Table 2 本方法 4 筆
表格設定(擷取紀錄原文):KITTI, trained on 00-06 and 09-10, tested on 07-08 to match LodoNet [32] (Wang et al., 2021c, Table 2)
trel (%),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 2)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LodoNet [32] | 1.86% | (Wang et al., 2021c, Table 2) |
| Ours (PWCLO-Net)本方法原文提出 | 0.61% | (Wang et al., 2021c, Table 2) |
Wang et al., 2021c · Table 3 本方法 4 筆
表格設定(擷取紀錄原文):KITTI, trained on 00-03 and 05-09, tested on 04 and 10 to match DeepPCO [27] (Wang et al., 2021c, Table 3)
trel (%),KITTI odometry · 04 (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 3)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| DeepPCO [27] | 2.63% | (Wang et al., 2021c, Table 3) |
| Ours (PWCLO-Net)本方法原文提出 | 0.73% | (Wang et al., 2021c, Table 3) |
Wang et al., 2021c · Table 4 本方法 4 筆
表格設定(擷取紀錄原文):KITTI, trained on 00-08, tested on 09 and 10 to match the unsupervised method of Cho et al. [3] (Wang et al., 2021c, Table 4)
trel (%),KITTI odometry · 09 (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 4)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Cho et al. [3] (unsupervised) | 4.87% | (Wang et al., 2021c, Table 4) |
| Ours (PWCLO-Net)本方法原文提出 | 0.73% | (Wang et al., 2021c, Table 4) |
來源
Wang et al., 2021c
(2021)PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 15905-15914
DOI 10.1109/cvpr46437.2021.01565arXiv 2012.00972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:PWCLO-Net (arXiv v2, accepted CVPR 2021 version) https://arxiv.org/abs/2012.00972
- 後續方法:EfficientLO-Net (IEEE TPAMI; not read) https://doi.org/10.1109/TPAMI.2022.3207015
- 程式碼釋出:IRMVLab/PWCLONet https://github.com/IRMVLab/PWCLONet
程式碼:https://github.com/IRMVLab/PWCLONet(授權:MIT (LICENSE file read))。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。