Geometrically stable sampling
作者以點對平面 ICP 線性化後的 6x6 共變異數矩陣(力與力矩項)分析幾何穩定性:特徵值偏小的特徵向量對應兩曲面可相互滑動的螺旋運動,並以條件數作為穩定度指標。取樣時先以稀疏隨機樣本估計重疊區的特徵向量,再依各點對每個特徵向量的約束量排序,貪婪補強目前最弱的方向,使條件數接近 1。合成溝槽平面的條件數由 66.1 降至 3.7(選取 30% 點),Forma Urbis Romae 碎片掃描中均勻取樣無法對齊的溝槽得以正確對齊。與長廊、隧道等幾何退化場景的關聯屬推論。
本頁內容
Detects pose uncertainty in ICP and selects samples that constrain potentially unstable transformations, addressing featureless-region failures.
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 | range-scanned meshes of Forma Urbis Romae fragments (scanner not named in the paper)、synthetic noisy meshes (incised plane and sphere) |
|---|---|
| 原文測試平台 | simulation (synthetic meshes)、static range scans of Forma Urbis Romae fragments |
| 狀態估計 | point-to-plane ICP linearized for small rotations; the 6x6 covariance matrix C of per-pair torque (p x n) and force (n) terms defines the normal equations; stability measured by the condition number of C (ratio of extreme eigenvalues, target close to 1), computed with points and normals of P after centring and scaling the points to unit mean distance |
| 資料關聯 | covariance sampling: estimate eigenvectors of C from several hundred random points in the overlap (overlap test by mesh-boundary check of closest points), build six binned lists of candidate points sorted by |x_k . v_i|, and greedily pick the next point from the list of the currently least-constrained eigenvector; closest points on Q then feed point-to-plane minimization; for Forma Urbis Romae meshes overlapping by 25%, several hundred random points sufficed to stabilize the eigenvector estimate |
| 時間表示 | 不適用 |
| 去畸變 | 不適用 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | triangle meshes or point sets with normals (normals averaged from adjacent faces or supplied externally) |
| 先驗資訊 | initial pose required (ICP) |
| 可輸出幾何 | sampling strategy and pose-uncertainty indication |
| 計算需求 | 5 s per ICP iteration with stable sampling vs 1.5 s with uniform sampling on a 400 MHz Pentium II (300,000-point meshes, 10% subsampled); about 5x uniform-sampling cost per iteration with the delayed overlap test and about 3x with seed-point crawling |
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 | 400MHz Pentium II | 執行運算平台 | 未標示 | 400 MHz | (Gelfand et al., 2003, Sec. 5) |
| 其他 | range scanner used for Forma Urbis Romae scans (model not named in the paper) | 資料集感測器 | Forma Urbis Romae | meshes of about 300,000 points | (Gelfand et al., 2003, Sec. 5; Acknowledgements) |
作者報告的優勢與限制
優勢
- condition number reduced from 66.1 to 3.7 (incised plane) and from 26.9 to 4.1 (incised sphere) with 30% of points (Figs. 4, 6)
- aligns grooves that uniform sampling misaligns; on the sphere it is the only method that finds the correct pose (Sec. 5, Figs. 5, 7)
- Forma Urbis Romae fragment converged in 25 iterations where uniform sampling failed; after global relaxation maximum residual 0.3 mm vs over 1 mm with uniform sampling (Sec. 5, Fig. 10)
- handles both translational and rotational instability, unlike normal-space sampling (Sec. 1, 3)
限制
- noisy areas can look like features and attract samples; smoothing helps only when features are larger than the noise, otherwise sampling fails (Sec. 5, Fig. 11)
- all eigenvectors are constrained equally; leverage of geometry outside the overlap is ignored (Sec. 6)
- pairwise only; stability of multi-scan global relaxation not addressed (Sec. 6)
- 3x to 5x slower per iteration than uniform sampling in the reported implementation (Sec. 4)
營建工程相關證據
原文未報告(與長廊、隧道等退化幾何的關聯屬推論)
原文驗證環境:模擬、受控實驗
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 9 個比較組,合計 18 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 5 組列在最後,並連到性能比較頁。
Tuna et al., 2025 · Table 3 本方法 4 筆
資料集與序列ANYmal simulation · pillar traverse
表格設定(擷取紀錄原文):Dynamic ANYmal simulation with Open3D SLAM in the loop; prior noise σt 0.05 m, σr 0.01 rad; EVO metrics (Tuna et al., 2025, Table 3)
APE translation µ(σ)[m]; σ 3.24,ANYmal simulation · pillar traverse
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 Tuna et al., 2025 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 |
|---|---|---|
| Gelfand et al.本方法 | 4.273 m | (Tuna et al., 2025, Table 3) |
Tuna et al., 2025 · Table 4 本方法 3 筆
資料集與序列ANYmal forest experiment · forest open field
表格設定(擷取紀錄原文):ANYmal forest to open-field run (107 m before revisit), leg-odometry prior, GNSS position ground truth; ICP-loop cost (Tuna et al., 2025, Table 4)
RTE µ(σ)[m], per 2 m traversed (Sec. V-B); σ 0.269,ANYmal forest experiment · forest open field
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 Tuna et al., 2025 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 |
|---|---|---|
| Gelfand et al.本方法 | 0.258 m | (Tuna et al., 2025, Table 4) |
Tuna et al., 2025 · Table 5 本方法 3 筆
資料集與序列ENWIDE (Ulmberg bicycle tunnel) · Ulmberg tunnel
表格設定(擷取紀錄原文):Ulmberg bicycle tunnel, handheld payload, COIN-LIO prior, one-directional degeneracy over more than 80% of the run; ICP registration cost (Tuna et al., 2025, Table 5)
RTE µ(σ)[m], per 1 m traversed; σ 0.075,ENWIDE (Ulmberg bicycle tunnel) · Ulmberg tunnel
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 Tuna et al., 2025 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 |
|---|---|---|
| Gelfand et al.本方法 | 0.073 m | (Tuna et al., 2025, Table 5) |
Gelfand et al., 2003 · Text Sec. 4 本方法 2 筆
資料集與序列原文未報告 · implementation variants
表格設定(擷取紀錄原文):Per-iteration cost relative to ICP with uniform sampling when meshes overlap by half their area (Gelfand et al., 2003, Text Sec. 4)
time per iteration relative to uniform-sampling ICP,原文未報告 · implementation variants
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 Gelfand et al., 2003 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 |
|---|---|---|
| stable sampling with delayed overlap test本方法原文提出 | 5 ratio | (Gelfand et al., 2003, Sec. 4) |
其他比較組
來源
Gelfand et al., 2003
(2003)Geometrically stable sampling for the ICP algorithmFourth International Conference on 3-D Digital Imaging and Modeling (3DIM 2003), pp. 260-267
同儕審查已出版已讀全文經典