Symmetric ICP
此文提出對稱化的 ICP 目標函數：以對應點兩側法向量的和作為誤差方向，並把旋轉拆成兩半，以相反方向分別作用於兩個表面。只要兩點與其法向量落在同一個局部二次曲面（2D 為圓弧）上，殘差即為零，因此比只在平面上為零的點對平面誤差保留更多沿表面滑動的自由度，卻不需估計曲率。作者以 Rodrigues 公式推導線性化，每次迭代仍只解一個線性最小平方問題，對應正確時結果是精確的。在 dragon 模型、部分重疊的 bunny 掃描與一組 TUM RGB-D 室內掃描上，每次迭代的誤差下降都快於點對點、點對平面、雙平面與二次近似法；bunny 收斂範圍實驗中，Levenberg-Marquardt 版本在 20 次迭代時收斂範圍最寬。論文沒有報告執行時間，也沒有使用 LiDAR 或大尺度場景資料。
本頁內容
Symmetric ICP measures the pair residual along the sum of both normals and splits the rotation between the two surfaces, so the error vanishes whenever the pair lies on a locally quadratic patch; one linear solve per iteration, exact for exact correspondences, with faster per-iteration convergence and a wider basin on object-scale scans.
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 scans and meshes (dragon model, Curless and Levoy 1996、bunny range scans bun000, bun090 and other overlapping pairs, Turk and Levoy 1994)", "RGB-D (Kinect scans from TUM RGB-D freiburg1_xyz)"] |
|---|---|
| 原文測試平台 | 未記錄 |
| 狀態估計 | E_symm = sum_i [(R p_i - R^-1 q_i + t) . (n_p,i + n_q,i)]^2: the rotation is split so that each surface is rotated by half the angle in opposite directions, with normals kept fixed (the rotated-normals variant E_symm-RN behaves similarly and is dropped for simplicity); linearized from the Rodrigues formula with a~ = a tan(theta) into one linear least-squares solve per iteration, interpreted as a Gauss-Newton step with Gibbs rotation parameters, exact when correspondences are exact; a Levenberg-Marquardt variant (LM-Symmetric) is also evaluated |
| 資料關聯 | both normals enter the residual (Sec. 3.2; if oriented normals are unavailable they are flipped to agree, footnote 1); in the convergence-basin test all variants sample points from both meshes and take Euclidean closest points on the other mesh (Sec. 4.4); the dragon self-alignment test uses no outlier rejection, while the partially overlapping bunny test and the basin test reject pairs with negative normal dot product and pairs farther than 2.5 sigma (sigma = 1.4826 x median distance) at each iteration (Sec. 4.3, 4.4) |
| 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 | 3D models / surfaces |
| 先驗資訊 | initial pose required |
| 可輸出幾何 | rigid transformation |
| 計算需求 | almost no added computation per iteration relative to point-to-plane, one linear least-squares solve per iteration; no runtime, timing table or hardware 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 |
|---|---|---|---|---|---|
| RGB-D 相機 | Kinect | 資料集感測器 | TUM RGB-D (freiburg1_xyz, timestamps 1305031104.030279 and 1305031108.503548) | 原文未報告 (paper notes warp and scanning noise from the Kinect sensor) | (Rusinkiewicz, 2019, Sec. 4.3) |
作者報告的優勢與限制
優勢
- per-iteration error decreases faster than point-to-point, point-to-plane, two-plane and quadratic objectives on the dragon model, a partially overlapping bunny pair and a TUM RGB-D office pair (Sec. 4.3, Fig. 4); linearization verified exact up to roundoff for exact correspondences (Sec. 4.2); LM-Symmetric gives the widest convergence basin at 20 iterations and comes within a few percent of two-plane at 500 iterations (Sec. 4.4); e.g. at 20 iterations with 60 deg rotation and 10% translation, 82% success for Symmetric vs 62% point-to-plane and 2% point-to-point (Fig. 5)
限制
- only object-scale data were used: dragon mesh, bunny range scans of Turk and Levoy (bun000 and bun090 overlap about 23% IOU) and one Kinect scan pair whose noise narrows the gap between variants; no LiDAR, large-scene or SLAM data (Sec. 4.3); plain E_symm has a basin comparable to the quadratic and LM point-to-plane variants and below two-plane at 500 iterations (Sec. 4.4); only a simple outlier-rejection scheme was tested, with robust l_p norms, denoising and descriptor matching left to future work (Sec. 4.3, 5); no timing numbers; still a local method needing an initial alignment (reviewer inference from Sec. 2 and 4.4)
營建工程相關證據
原文未報告（僅以 dragon 模型、bunny 掃描與一組 TUM RGB-D 室內辦公室掃描測試，未涉及營建場域或 LiDAR 資料）
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 1 個比較組，合計 36 筆紀錄。
Rusinkiewicz, 2019 · Fig. 5 本方法 36 筆
指標percentage of successful ICP trials
表格設定（擷取紀錄原文，這些數值分屬表中不同部分）：(Rusinkiewicz, 2019, Fig. 5)
- Numerals printed in Fig. 5 heatmap cells: % of 1000 random initial transforms (given rotation about a random axis, translation as fraction of mesh size), averaged over all bunny scan pairs with IOU > 20%, that end within 1% of mesh size of ground truth; 4 of 24 cells per variant and iteration budget transcribed
- Same setting as other Fig. 5 rows
percentage of successful ICP trials，bunny range scans (Turk and Levoy 1994) · 20 iterations; rotation 20 deg, translation 10%
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 Rusinkiewicz, 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(Rusinkiewicz, 2019, Fig. 5)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Point-to-point | 22% | (Rusinkiewicz, 2019, Fig. 5) |
| Quadratic (Mitra et al. 2004, on-demand) | 98% | (Rusinkiewicz, 2019, Fig. 5) |
| Point-to-plane | 99% | (Rusinkiewicz, 2019, Fig. 5) |
| Two-plane | 99% | (Rusinkiewicz, 2019, Fig. 5) |
| Symmetric-RN本方法原文提出 | 99% | (Rusinkiewicz, 2019, Fig. 5) |
| Symmetric本方法原文提出 | 99% | (Rusinkiewicz, 2019, Fig. 5) |
| LM-Point-to-plane (Fitzgibbon 2001) | 99% | (Rusinkiewicz, 2019, Fig. 5) |
| LM-Symmetric本方法原文提出 | 99% | (Rusinkiewicz, 2019, Fig. 5) |
來源
Rusinkiewicz, 2019
（2019）A symmetric objective function for ICPACM Transactions on Graphics, 38(4):1-7
同儕審查已出版已讀全文近十年