Glass reflection removal
地面雷射掃描遇到玻璃時，同一雷射脈衝可能同時產生玻璃點、穿透點，以及經玻璃反射而落在玻璃後方的虛像點。本法利用 RIEGL VZ-400 的多回波特性，把單位球面切成約 3×3 個脈衝的面片並計算投影點數，以兩成分高斯混合模型與 EM 分出一般與玻璃面片，再以 RANSAC 擬合最靠近掃描儀的主要玻璃平面，並依距離計算可靠度。玻璃平面後方的點若經 Householder 鏡射後能找到位置相近且 FPFH 特徵相似的真實點，就判為虛像點，最後以資料項加鄰點平滑項組成的標記能量函數，用 ICM 求解後移除虛像點。以六個含玻璃的戶外場景點雲作定性驗證，每個模型處理時間約 55 至 135 s。
本頁內容
Removes glass-reflection ghost points from single TLS scans: echo counts per sphere patch of about 3x3 pulses are classified by a two-component GMM, the dominant glass plane is fitted by RANSAC and weighted by distance-based reliability, and points behind it are labelled virtual when a Householder-mirrored real counterpart with similar FPFH exists, refined by a labelling energy with data and neighbour smoothness terms solved with ICM; validated qualitatively on six outdoor RIEGL VZ-400 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 | RIEGL VZ-400 terrestrial laser scanner with multiple echo returns, angular resolution 0.06° x 0.06° |
|---|---|
| 原文測試平台 | static TLS |
| 狀態估計 | two-component Gaussian mixture (EM) on per-patch point counts; RANSAC glass-plane fit with distance-weighted reliability; per-point score from reflection-symmetry distance and FPFH Hellinger similarity; binary labelling by an energy with data and neighbour smoothness terms (MRF-type; the paper does not use the term) minimised with ICM |
| 資料關聯 | k-d tree nearest real point to the Householder-mirrored position of each candidate behind the glass plane; FPFH on 50 nearest neighbours; 48-neighbour smoothness term limited to 0.1% of the bounding-box diagonal |
| 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 | 不適用 |
| 先驗資訊 | multi-echo information |
| 可輸出幾何 | point cloud with virtual reflection points removed |
| 計算需求 | Intel i7-4790K (4.38 GHz as written); 54.7 to 134.8 s per model of 4.9 to 9.7 million points, with descriptor computation taking more than half of the total time in most cases (5 of 6 models in Table 1) |
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 |
|---|---|---|---|---|---|
| 地面雷射掃描儀（TLS） | RIEGL VZ-400歸入：Riegl VZ-400 | 方法輸入 | 未標示 | multiple echo returns; angular resolution 0.06° azimuthal and 0.06° polar; about 5 to 6 million points per model | (Yun & Sim, 2018, Sec. 2, 5) |
| 運算硬體 | Intel i7-4790k | 執行運算平台 | 未標示 | 4.38 GHz as written | (Yun & Sim, 2018, Sec. 5; Table 1) |
作者報告的優勢與限制
優勢
- Faithful glass-region estimation and removal of virtual points on outdoor TLS scenes (abstract, Sec. 5, Fig. 9)
- Presented by the authors as the first reflection-removal algorithm for large-scale 3D point clouds (Sec. 1)
- Uses geometry and echo counts only, so no camera is needed (Sec. 7)
限制
- Assumes a single dominant glass plane per scan (Sec. 7)
- No quantitative evaluation because reflection-free ground truth is hard to obtain (Sec. 7)
- Virtual points remain when their real counterparts are occluded or glass patches get low reliability, and real trees crossing the extended glass plane can be removed (Sec. 5)
- β1 and β2 are set empirically per model (Sec. 5)
- Tested on static TLS outdoor scenes; applicability to mobile SLAM scans not shown (inference)
營建工程相關證據
原文未報告。作者以建物玻璃帷幕與車輛玻璃造成的反射虛像為動機（Sec. 1, Sec. 2），測試對象為六個戶外建物與庭園場景，未涉及施工工地，也沒有定量評估。玻璃帷幕與窗戶在既有建物與完工階段常見，虛像點可能干擾竣工比對與 Scan-to-BIM（推論）。
原文驗證環境：已完工建築
報告的性能數據
性能數據仍在分批查證，目前尚未收錄此方法的報告值。
來源
Yun & Sim, 2018
（2018）Reflection Removal for Large-Scale 3D Point Clouds2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4597-4605
同儕審查已出版已讀全文近十年