CELLO-3D
作者先檢視既有封閉形式共變異數估計在 3D 資料上的限制，再以資料驅動方式學習 ICP 配準的共變異數。訓練與評估使用超過五百萬次配準、1020 組真實點雲對，涵蓋結構化與非結構化、室內與室外環境。
本頁內容
Learns ICP registration covariance from large-scale real registrations after showing the limits of closed-form estimators on 3D data.
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 point clouds from the 'Challenging data sets for point cloud registration algorithms' (Pomerleau et al. 2012); the sensor is not named in this paper |
|---|---|
| 原文測試平台 | offline evaluation on recorded public datasets (no platform named in this paper) |
| 狀態估計 | CELLO-style kernel-weighted average of training covariances; an upper-triangular distance metric is learned by SGD with a determinant-plus-trace loss; descriptors per cell of a 4x4x4 grid (25 m x 25 m x 10 m) over the overlap region hold planarity, cylindricality and a 9-bin normal histogram; training covariances come from 5000 ICP samples per pair filtered with DBSCAN |
| 資料關聯 | point-to-plane ICP configured in the framework of Pomerleau et al. [24]: maximum-density and random subsampling filters, k-d tree 3-nearest-neighbour matching, trimmed-distance outlier filter keeping the closest 70%, at most 80 iterations |
| 時間表示 | 不適用 |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | 不適用 |
| 全域最佳化 | none |
| 地圖表示 | 不適用 |
| 先驗資訊 | none |
| 可輸出幾何 | registration covariance |
| 計算需求 | training covariances from about 5,100,000 registrations on 1020 pairs computed on Compute Canada clusters in about 5 CPU-years; online inference time not reported |
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 | Compute Canada computing clusters | 執行運算平台 | 未標示 | about 5 CPU-years for about 5,100,000 training registrations (offline data generation, not online runtime) | (Landry et al., 2019, Sec. V-A) |
作者報告的優勢與限制
優勢
- Lower average KL divergence than the mean-covariance baseline on all seven test groups, with the largest gains indoors, e.g. Apartment 26.6 vs 34.1 (Table II)
- Censi's closed-form estimate had average KL divergences of 2.25e6 to 2.06e8 because it was orders of magnitude too small (Table II, Sec. VI-A)
- Average Mahalanobis distance of final odometry poses between 0.405 and 2.50 over 100 trajectories per dataset, which the authors judge consistent overall (Table III, Sec. VI-B)
限制
- Gains over the baseline are modest in self-similar environments (Wood, Gazebo), where the learned weights become nearly uniform (Sec. VI-A)
- Estimates are pessimistic for Stairs and Apartment (average DM below 1.5) (Sec. VI-B)
- Treating ICP results as normally distributed is error-prone in SE(3) because observed distributions are mainly multimodal, and descriptor quality is critical (Sec. VII)
- Data augmentation only rotates about the z axis (2.5D), and each test group was trained on another dataset of the same environment type (Sec. V-A, VI, Table II)
營建工程相關證據
原文未報告
原文驗證環境：公開基準、已完工建築、獨立參考量測
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 3 個比較組，合計 58 筆紀錄。
Landry et al., 2019 · Table III 本方法 35 筆
表格設定（擷取紀錄原文）：ICP odometry over whole sequence, initial guesses sampled with a = 0.05; final pose error and Mahalanobis distance DM against ground truth; DM averaged over 100 trajectories (Landry et al., 2019, Table III)
final translation error ||u||，Challenging data sets (ETH) · Apartment (22 m)
這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Landry et al., 2019 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| CELLO-3D (ICP odometry)本方法原文提出 | 0.115 m | (Landry et al., 2019, Table III) |
Brossard et al., 2020 · Table 2 本方法 16 筆
表格設定（擷取紀錄原文）：Trajectory consistency: Mahalanobis distance of compounded ICP trajectories to ground truth, averaged over 40 initial trajectories per sequence; target 1, below 1 pessimistic; CELLO-3D reproduced from Landry et al. with a slightly different ICP setting (Brossard et al., 2020, Table 2)
Mah. dist. trans.，Challenging data sets for point cloud registration (Pomerleau et al. 2012) · Apartment
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 Brossard et al., 2020 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(Brossard et al., 2020, Table 2)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| CELLO-3D本方法 | 0.2 | (Brossard et al., 2020, Table 2) |
| ini.+ICP (fusion without cross-covariance) | 3.5 | (Brossard et al., 2020, Table 2) |
| proposed (full ML covariance, Eq. 15)原文提出 | 2.3 | (Brossard et al., 2020, Table 2) |
Landry et al., 2019 · Table II 本方法 7 筆
指標Avg. KL divergence
表格設定（擷取紀錄原文，這些數值分屬表中不同部分）：(Landry et al., 2019, Table II)
- Average KL divergence between sampled and predicted ICP covariance per test group; trained on the named other group
- Average KL divergence per test group
Avg. KL divergence，Challenging data sets (ETH) · Apartment (trained on Haupt. and Stairs; 1190 pairs)
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 Landry 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(Landry et al., 2019, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Baseline (mean training covariance) | 34.1 | (Landry et al., 2019, Table II) |
| Ours (CELLO-3D)本方法原文提出 | 26.6 | (Landry et al., 2019, Table II) |
| Censi | 91900000 | (Landry et al., 2019, Table II) |
來源
Landry et al., 2019
（2019）CELLO-3D: Estimating the Covariance of ICP in the Real World2019 International Conference on Robotics and Automation (ICRA), pp. 8190-8196
DOI 10.1109/icra.2019.8793516arXiv 1810.01470
同儕審查已出版已讀全文近十年
相關版本
- 預印本：CELLO-3D: Estimating the Covariance of ICP in the Real World https://arxiv.org/abs/1810.01470