Screened Poisson Surface Reconstruction
此版本在原泊松重建中加入點位置的軟約束(screening term),使重建等值面更貼近輸入點,以減輕原方法的過度平滑;因約束只定義在稀疏點集上,線性系統的稀疏結構不變,仍可用多重網格(multigrid)求解,並透過演算法改良使時間複雜度對點數呈線性。作者另支援 Neumann 邊界條件,讓缺資料區的曲面可延伸到定義域邊界,而非強制封閉。
本頁內容
Adds sparse point-interpolation (screening) constraints to Poisson reconstruction to reduce over-smoothing, keeps a multigrid-solvable sparse system with linear-time complexity, and supports Neumann boundaries.
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 | 不適用 |
| 資料關聯 | 不適用 |
| 時間表示 | 不適用 |
| 去畸變 | 不適用 |
| 迴圈閉合 | 不適用 |
| 全域最佳化 | global screened Poisson solve with multigrid on an octree |
| 地圖表示 | implicit function on an adaptive octree with point-value (screening) constraints |
| 先驗資訊 | oriented normals required |
| 可輸出幾何 | triangle mesh isosurface; Dirichlet (closed) or Neumann (open to domain boundary) boundary behavior |
| 計算需求 | offline, CPU; C++ with OpenMP multithreading and conjugate-gradient relaxation at each multigrid level (Sec. 5.4); Table I timings on a laptop with a quad-core Intel Core i7 and 8 GB RAM (Sec. 6.2) |
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 |
|---|---|---|---|---|---|
| 運算硬體 | Intel Core i7 | 執行運算平台 | 未標示 | quad-core; laptop with 8 GB RAM | (Kazhdan & Hoppe, 2013, Sec. 6.2) |
作者報告的優勢與限制
優勢
- ["Reconstructions better capture input data at similar complexity and time as unscreened Poisson (Sec. 1, Fig. 1).", "Solver complexity reduced to linear in number of points (abstract).", "One-sided RMS error from held-out validation points is always lower than original Poisson and Wavelet and comparable to or lower than SSD on real scans
- the same ordering holds on clean uniformly sampled data (Sec. 6.1, Figs. 4b and 5b).", "Faster than original Poisson at depths 9 to 11, e.g. David at depth 10: 182 s versus 412 s, while SSD needed 19,158 s (Table I).", "Authors state that with screening off (alpha = 0) the new solver is 2 to 3 times faster than the original Poisson implementation, of which a factor 1.1 to 1.6 comes from multithreading (Sec. 6.2)."]
限制
- ["With misaligned input scans, the screened reconstruction produces a pock-marked surface that undulates between scans (Sec. 6.3).", "For noisy data the screening weight must be reduced (Sec. 6.3).", "Neumann boundaries bias surfaces to cross the domain boundary orthogonally (Sec. 4.4, Fig. 3).", "SSD extrapolates better into regions of missing data (e.g. the Anchor model's cylindrical hole), giving SSD the smallest distance to ground truth in the Berger et al. benchmark (Sec. 6.1, Fig. 6
- Sec. 7).", "Memory is higher than original Poisson, e.g. David at depth 10: 2194 MB versus 1498 MB (Table I).", "Point-to-surface RMS can favour the screened result even when misaligned scans make it visually worse (Sec. 6.3)."]
營建工程相關證據
未在營建資料驗證,測試資料為雕像與機械零件等物件掃描及合成掃描。對 SLAM 點雲而言,作者報告的「掃描未對齊時表面起伏」代表軌跡或迴圈誤差會直接轉成網格表面瑕疵(推論);作者也建議雜訊大時降低 screening 權重 α,而資料缺漏處的外插能力不如 SSD。
原文驗證環境:公開基準、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 28 筆紀錄。
Kazhdan & Hoppe, 2013 · Table I 本方法 24 筆
表格設定(擷取紀錄原文):Wall-clock time and memory for Neptune and David at depths 8 to 11; screening weight alpha = 4, Neumann boundaries, samples-per-node 1; bracketed values are the new solver with alpha = 0; dagger: SSD at David depth 11 reports CPU user time because memory exceeded RAM (Kazhdan & Hoppe, 2013, Table I)
Time in seconds,Neptune (Aim@Shape) · depth 8
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 Kazhdan & Hoppe, 2013 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(Kazhdan & Hoppe, 2013, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Poisson硬體:laptop, quad-core Intel Core i7, 8 GB RAM | 10 s | (Kazhdan & Hoppe, 2013, Table I) |
| Wavelet硬體:laptop, quad-core Intel Core i7, 8 GB RAM | 3 s | (Kazhdan & Hoppe, 2013, Table I) |
| SSD硬體:laptop, quad-core Intel Core i7, 8 GB RAM | 275 s | (Kazhdan & Hoppe, 2013, Table I) |
| Screened本方法原文提出硬體:laptop, quad-core Intel Core i7, 8 GB RAM | 14 s | (Kazhdan & Hoppe, 2013, Table I) |
Kazhdan & Hoppe, 2013 · Text Sec. 1 本方法 4 筆
資料集與序列David (Digital Michelangelo, 11.4M-point subset) · David head, depth 10
表格設定(擷取紀錄原文):Subset of 11.4M points from the David scan, octree depth 10 (effective 1024^3), timings without parallelization; unscreened timing uses the new implementation with screening weight 0 (Kazhdan & Hoppe, 2013, Text Sec. 1)
processing time without parallelization,David (Digital Michelangelo, 11.4M-point subset) · David head, depth 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 Kazhdan & Hoppe, 2013 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(Kazhdan & Hoppe, 2013, Text Sec. 1)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| traditional Poisson (new implementation, screening weight 0)本方法原文提出 | 230 s | (Kazhdan & Hoppe, 2013, Sec. 1, footnote 1, Fig. 1) |
| screened Poisson本方法原文提出 | 272 s | (Kazhdan & Hoppe, 2013, Sec. 1, Fig. 1) |
來源
Kazhdan & Hoppe, 2013
(2013)Screened poisson surface reconstructionACM Transactions on Graphics, 32(3):1-13
DOI 10.1145/2487228.2487237程式碼
同儕審查已出版已讀全文經典查證後修正
相關版本
- 前身方法:Poisson Surface Reconstruction (Kazhdan, Bolitho and Hoppe, SGP 2006), the unscreened predecessor method; not a conference version of this article 10.2312/SGP/SGP06/061-070
程式碼:https://github.com/mkazhdan/PoissonRecon(授權:MIT (repository LICENSE))。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。