Volumetric range-image integration (TSDF origin; VRIP)
作者將每張已對齊的距離影像(range image)沿感測器視線轉成有號距離函數與權重,逐一加權累加到體素格網中,最後擷取零等值面成為三角網格;在特定假設下,此等值面在最小平方意義上最佳。體素另外標記為空、未觀測或近表面三種狀態,藉由空間雕刻(space carving)在空與未觀測區域的交界補面,產生無孔洞模型。
本頁內容
Fuses aligned range images into a cumulative weighted signed distance field along sensor lines of sight and extracts the zero isosurface, with space carving used to fill holes.
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 | ["Cyberware 3030 MS laser-stripe optical triangulation scanner (traditional triangulation or spacetime analysis)"] |
|---|---|
| 原文測試平台 | ["laser-stripe optical triangulation scanner with the object translated through the laser plane (the Fig. 5 caption calls the Cyberware a translating sensor)、objects repositioned between scans by a motion control platform"] |
| 狀態估計 | 不適用 (range images are assumed pre-aligned) |
| 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 | none |
| 地圖表示 | cumulative weighted signed distance function on a voxel grid (run-length encoded) with empty/unseen/near-surface voxel states |
| 先驗資訊 | aligned range images |
| 可輸出幾何 | watertight triangle mesh extracted as the zero isosurface; hole filling by tessellating between empty and unseen regions |
| 計算需求 | offline; 250 MHz MIPS R4400: observed-surface integration under 1 h (47 to 56 min) and space carving with hole filling 3 to 5 h (197 to 257 min); up to 12 million input vertices and grids up to 160 million voxels |
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 |
|---|---|---|---|---|---|
| 載具平台 | motion control platform (model not reported) | 方法輸入 | 未標示 | used to reposition objects automatically between scans | (Curless & Levoy, 1996, Sec. 6) |
| 運算硬體 | MIPS R4400 | 執行運算平台 | 未標示 | 250 MHz processor | (Curless & Levoy, 1996, Sec. 6) |
| 其他 | Cyberware 3030 MS | 方法輸入 | 未標示 | laser stripe optical triangulation scanner; laser sheet from a cylindrical lens with an off-axis CCD; object translates through the laser plane; modified to permit spacetime triangulation | (Curless & Levoy, 1996, Sec. 5.1, Fig. 1(b), Fig. 12 caption) |
作者報告的優勢與限制
優勢
- ["Integrates many range images (up to 70) into seamless models of up to 2.6 million triangles (abstract).", "Incremental, order-independent updates and a least-squares-optimal isosurface under stated assumptions (orthographic sensor, independent errors along lines of sight) (Sec. 3, Appendix A sketch).", "RMS distance between original range points and the reconstructed dragon and Buddha surfaces is about 0.1 mm, roughly the scanner accuracy (Sec. 6).", "Behaves robustly on a 1.6 mm drill bit scanned from 12 orientations, where the zippering method fails catastrophically (Sec. 6, Fig. 9).", "Run-length encoding gives typical memory savings of 10:1 to 20:1 (Sec. 5.2.1)."]
限制
- ["Difficulty bridging sharp corners when no scan spans both adjoining surfaces (Sec. 7).", "Thin surfaces are problematic: distance ramps extending behind surfaces interfere across opposite sides, causing thickening of thin surfaces and rounding of sharp corners (Sec. 7).", "Hole-fill surfaces are not observed geometry
- without extra backdrop carving they create extraneous tessellations, and their unseen to empty transition produces aliasing that needs separate filtering (Sec. 4, Fig. 11).", "Optical scanning limits: only external surfaces are seen, shiny, dark or bright surfaces cause errors, and the authors often paint objects flat gray to reduce these effects (the Buddha was painted matte gray, the drill bit white) (Sec. 7, Figs. 9 and 12).", "Range images must be pre-aligned and weights follow optical-triangulation uncertainty (Sec. 3)."]
營建工程相關證據
原研究為小型物件掃描(龍、20 cm 高的快樂佛像、1.6 mm 鑽頭),未在營建場域驗證;作者僅在 Sec. 7 提到未來希望應用到地形與建築場景。作者自述的薄面增厚與尖角圓化限制,直接對應牆板、鋼構翼板與構件邊緣等營建幾何的尺寸風險;孔洞填補產生的面並非量測資料,用於營建成果時需與觀測面分開標示(推論)。
原文驗證環境: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 5 個比較組,合計 30 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 1 組列在最後,並連到性能比較頁。
Curless & Levoy, 1996 · Fig. 8 本方法 12 筆
表格設定(擷取紀錄原文):Reconstruction statistics with and without space carving and hole filling; Dragon voxel 0.35 mm (712x501x322), Buddha voxel 0.25 mm (407x957x407) (Curless & Levoy, 1996, Fig. 8)
Exec. time (min),Dragon (authors' scans) · 61 scans, 15 M input triangles
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 Curless & Levoy, 1996 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 |
|---|---|---|
| Dragon (volumetric integration, no hole filling)本方法原文提出硬體:250 MHz MIPS R4400 | 56 min | (Curless & Levoy, 1996, Fig. 8) |
Wang et al., 2023a · Table 1 本方法 8 筆
表格設定(擷取紀錄原文):Averages over scenes; meshes culled with the authors' new culling strategy for all methods; TSDF-Fusion uses Co-SLAM poses; iMAP* is the NICE-SLAM re-implementation (Wang et al., 2023a, Table 1)
Depth L1 (cm),Replica (8 synthetic scenes) · average over scenes
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 Wang et al., 2023a 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(Wang et al., 2023a, Table 1)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| TSDF-Fusion本方法 | 6.36 cm | (Wang et al., 2023a, Table 1) |
| iMAP | 4.64 cm | (Wang et al., 2023a, Table 1) |
| NICE-SLAM | 1.9 cm | (Wang et al., 2023a, Table 1) |
| Co-SLAM (Ours)原文提出 | 1.51 cm | (Wang et al., 2023a, Table 1) |
Zhu et al., 2022a · Table 1 本方法 5 筆
資料集與序列Replica · average of 8 scenes
表格設定(擷取紀錄原文):Replica, average over 8 scenes and 5 runs; 3D metrics computed after removing regions outside every camera frustum; TSDF-Fusion uses NICE-SLAM poses at 256^3 voxels; iMAP* is the authors' re-implementation (Zhu et al., 2022a, Table 1)
Mem. (MB),Replica · average of 8 scenes
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 Zhu et al., 2022a 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(Zhu et al., 2022a, Table 1)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| TSDF-Fusion [11] (with NICE-SLAM poses)本方法 | 67.1 MB | (Zhu et al., 2022a, Table 1) |
| iMAP* [47] (re-implementation) | 1.04 MB | (Zhu et al., 2022a, Table 1) |
| DI-Fusion [16] | 3.78 MB | (Zhu et al., 2022a, Table 1) |
| NICE-SLAM原文提出 | 12.02 MB | (Zhu et al., 2022a, Table 1) |
Kazhdan et al., 2006 · Table 2 本方法 3 筆
資料集與序列Stanford Bunny · raw data, 362,000 points (Poisson at depth 9)
表格設定(擷取紀錄原文):Stanford Bunny raw data (362,000 points from ten range images), processed to fit each algorithm's input format; Poisson reconstructed at octree depth 9, resolution settings of the other seven methods not reported (VRIP used the registered scans with confidence values); running time in seconds, peak memory in MB, output triangles; hardware not reported (Kazhdan et al., 2006, Table 2)
Time (s),Stanford Bunny · raw data, 362,000 points (Poisson at depth 9)
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 et al., 2006 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 et al., 2006, Table 2)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Power Crust | 380 s | (Kazhdan et al., 2006, Table 2) |
| Robust Cocone | 892 s | (Kazhdan et al., 2006, Table 2) |
| FastRBF | 4919 s | (Kazhdan et al., 2006, Table 2) |
| MPU | 28 s | (Kazhdan et al., 2006, Table 2) |
| Hoppe et al 1992 | 70 s | (Kazhdan et al., 2006, Table 2) |
| VRIP本方法 | 86 s | (Kazhdan et al., 2006, Table 2) |
| FFT | 125 s | (Kazhdan et al., 2006, Table 2) |
| Poisson原文提出 | 263 s | (Kazhdan et al., 2006, Table 2) |
其他比較組
列出其餘 1 個比較組
來源
Curless & Levoy, 1996
(1996)A volumetric method for building complex models from range imagesProceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH '96), pp. 303-312
同儕審查已出版已讀全文經典