Sparse ICP
作者指出一般 ICP 依賴修剪或重新加權對應點的經驗法則來處理離群值與部分重疊,這些法則不穩定且難以調整。論文將配準目標改為對每組對應點的殘差向量施加 p 介於 0 與 1 之間的 ℓp 範數(群組稀疏),保留最近點搜尋步驟,並以交替方向乘子法(ADMM)與收縮運算子求解剛體轉換,點對點與線性化點對平面兩種版本皆可使用,實驗多採 p = 0.4。在虛擬掃描的貓頭鷹模型上,其配準 RMSE 為 4.8e-4,優於 ℓ1-ICP 的 1.6e-2 與使用距離門檻剔除的傳統 ICP。作者也指出 p 越小收斂越慢,而目標幾何含大量離群值或初始位置相距過遠時,最近點步驟仍會使結果落入錯誤的局部極小值。
本頁內容
Replaces heuristic correspondence pruning with sparsity-inducing norms in ICP to handle outliers and missing 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 | 未記錄 |
|---|---|
| 原文測試平台 | 未記錄 |
| 狀態估計 | l_p norm (p in [0, 1], p = 0.4 by default) of per-correspondence residual vectors (group sparsity) minimized with ADMM: a shrinkage step on auxiliary variables, a classical least-squares rigid fit, and a multiplier update; point-to-point and linearized point-to-plane versions (Sec. 4 to 7) |
| 資料關聯 | closest point on the target via an l2 kd-tree; unchanged by the l_p metric because |r|^p is monotone (Sec. 5.1) |
| 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 scans / geometric data sets |
| 先驗資訊 | initial pose required |
| 可輸出幾何 | rigid transformation |
| 計算需求 | 原文未報告 (C++ implementation released; no runtime or hardware reported) |
使用設備
尚未收錄此方法的設備紀錄;設備資料仍在分批查證,沒有紀錄不代表原文未使用任何設備。
作者報告的優勢與限制
優勢
- ["heuristic-free outlier handling with a single free parameter p (Sec. 8)", "on the virtually scanned owl model, RMSE 4.8e-4 versus 1.6e-2 for l1-ICP and 2.9e-2 to 4.1e-1 for thresholded least-squares ICP (Fig. 4)", "ADMM is stable and faster than iterative reweighting for point-to-plane, where reweighting is ill-conditioned (Sec. 7, Fig. 10)", "selects inliers automatically, unlike trimmed ICP which needs the inlier ratio (Sec. 7, Fig. 12)", "handles whole-in-part registration where Tukey and fair weights fail (Fig. 7)"]
限制
- ["the closest-point step is still affected by outliers, so performance degrades with many outliers in the target geometry (Sec. 7 Limitations, Fig. 9)", "local method: converges to a wrong minimum when source and target start far apart (Fig. 9)", "convergence slows as p decreases
- p = 0.4 chosen empirically as a trade-off (Sec. 7, Fig. 11)", "evaluation is a limited set of illustrative experiments rather than an exhaustive comparison (Sec. 7)"]
營建工程相關證據
原文未報告(實驗對象為雕像與物件的虛擬掃描、合成雜訊掃描及消費級深度相機掃描,未涉及建物或工地)
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 2 個比較組,合計 13 筆紀錄。
Zhou et al., 2016 · Table 3 本方法 12 筆
表格設定(擷取紀錄原文):Timing of local refinement methods and FGR on the same five synthetic models as Table 2 (same point counts; FGR column identical to Table 2); the caption does not state how per-model times were aggregated; single thread (Zhou et al., 2016, Table 3)
running time per pairwise registration (s),Synthetic range images · Bimba (9,416 points avg)
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 Zhou et al., 2016 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(Zhou et al., 2016, Table 3)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| PCL ICP point-to-point硬體:Intel Core i7-5960X 3.00 GHz, single thread | 0.73 s | (Zhou et al., 2016, Table 3) |
| PCL ICP point-to-plane硬體:Intel Core i7-5960X 3.00 GHz, single thread | 0.31 s | (Zhou et al., 2016, Table 3) |
| Sparse ICP point-to-point [5]本方法硬體:Intel Core i7-5960X 3.00 GHz, single thread | 3.1 s | (Zhou et al., 2016, Table 3) |
| Sparse ICP point-to-plane [5]本方法硬體:Intel Core i7-5960X 3.00 GHz, single thread | 11.8 s | (Zhou et al., 2016, Table 3) |
| Our approach (FGR)原文提出硬體:Intel Core i7-5960X 3.00 GHz, single thread | 0.13 s | (Zhou et al., 2016, Table 3) |
Bouaziz et al., 2013 · Fig. 4 本方法 1 筆
指標e, RMSE w.r.t. ground-truth alignment, panel (f)
資料集與序列virtually scanned 'owl' statue model (lgg.epfl.ch/statues) · owl
表格設定(擷取紀錄原文):Virtually scanned 'owl' model registered to its ground truth; e = RMSE of registered point locations with respect to the ground-truth alignment; values are the numbers printed under each panel of Fig. 4 (not read from an axis); distance thresholds are percentages of the bounding-box diagonal (Bouaziz et al., 2013, Fig. 4)
e, RMSE w.r.t. ground-truth alignment, panel (f),virtually scanned 'owl' statue model (lgg.epfl.ch/statues) · owl
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 Bouaziz et al., 2013 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 |
|---|---|---|
| lp-ICP, p = 0.4 (proposed)本方法原文提出 | 0.00048 | (Bouaziz et al., 2013, Fig. 4 and caption; Sec. 7) |
來源
Bouaziz et al., 2013
(2013)Sparse Iterative Closest PointComputer Graphics Forum, 32(5):113-123
同儕審查已出版已讀全文經典查證後修正
程式碼:https://github.com/opengp/sparseicp(授權:MPL-2.0 (stated in README; no LICENSE file))。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。