Point-based fusion
此系統全程只用一個扁平的點（surfel）清單表示場景，每點存位置、法向量、半徑、信心計數與時間戳，不建立體素或其他空間資料結構。每個影格先以三層階層式稠密 ICP 將深度圖對齊到由模型點渲染出的深度圖來估計 6DoF 位姿，再把模型點渲染成超取樣索引圖做投影式資料關聯；對應點依距影像中心遠近的高斯信心加權平均融合，信心累積到門檻 10 才由不穩定轉為穩定。系統另移除違反自由空間的點，並從 ICP 找不到對應的像素出發做區域成長，把移動物體整體標為動態並排除於位姿估計之外。
本頁內容
A flat, unstructured list of surfels (position, normal, radius, confidence, timestamp) is updated in the graphics pipeline: frame-to-model hierarchical ICP gives the pose, a supersampled index map provides projective association, confidence-weighted averaging fuses points, free-space violations are removed, and region growing from ICP outliers segments dynamic objects.
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 | ["RGB-D camera (Microsoft Kinect, near mode, 640x480 depth)", "time-of-flight camera (PMD CamBoard, 200x200, per-pixel amplitude used for confidence)"] |
|---|---|
| 原文測試平台 | ["stationary Kinect with objects on a turntable (Flowerpot, Teapot、sequences recorded by Nguyen et al.)", "Kinect (near mode stated for most scenes) in the indoor Large Office (two rooms and connecting corridors, about 10 m x 6 m x 2.5 m), Moving Person and Ballgame scenes、camera carrier and motion are not described per scene", "PMD CamBoard ToF camera (PMD scene、carrier not stated)", "synthetic depth maps from a virtual camera rotating around the Sim scene"] |
| 狀態估計 | frame-to-model dense hierarchical (pyramid) ICP against the rendered model point map, estimating a 6-DoF pose before fusion |
| 資料關聯 | projective data association by rendering the global point model as an index map |
| 時間表示 | discrete poses |
| 去畸變 | 不適用 (each depth map is registered with a single 6DoF camera pose; the paper does not discuss shutter type or intra-frame motion) |
| 迴圈閉合 | none (authors state sensor drift is not tackled; loop closure is future work, Sec. 8) |
| 全域最佳化 | none |
| 地圖表示 | flat list of points/surfels with position, normal, radius and confidence counter; unstable-to-stable status; no spatial data structure |
| 先驗資訊 | none |
| 可輸出幾何 | fused point (surfel) model with position, normal, radius, confidence and timestamp, optionally the last RGB sample per point; visualised by opaque surface splatting; no mesh extraction |
| 計算需求 | GPU (Intel i7 8-core CPU, NVIDIA GTX 680); 640x480 input (200x200 for the PMD scene); per-frame ICP about 11-22 ms, dynamic segmentation about 0.7-3.2 ms and fusion about 3-18 ms; processed rates 15 to 27 fps (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 |
|---|---|---|---|---|---|
| RGB-D 相機 | Microsoft Kinect | 方法輸入 | 未標示 | near mode; 640x480 depth input; up to 307,200 points per frame; 30 fps input | (Keller et al., 2013, Sec. 3, Sec. 7, Table 1) |
| 載具平台 | turntable (model not reported) | 資料集感測器 | Flowerpot and Teapot (Nguyen et al.) | objects rotated in front of a stationary Kinect | (Keller et al., 2013, Sec. 7) |
| 運算硬體 | Intel i7 8-core CPU | 執行運算平台 | 未標示 | 8-core | (Keller et al., 2013, Table 1 caption) |
| 運算硬體 | NVidia GTX 680 | 執行運算平台 | 未標示 | GPU | (Keller et al., 2013, Table 1 caption) |
| 其他 | PMD CamBoard | 方法輸入 | 未標示 | time-of-flight camera; 200x200 frames; per-pixel amplitude used in sample confidence; 27 fps input | (Keller et al., 2013, Sec. 3, Sec. 7, Table 1, Fig. 10) |
| 其他 | Vicon | 參考或真值量測 | Flowerpot and Teapot (Nguyen et al.) | motion capture ground truth for Kinect poses | (Keller et al., 2013, Sec. 7, Fig. 5) |
作者報告的優勢與限制
優勢
- ["Speed and memory efficiency from the point representation
- robustness to dynamic scenes (abstract).", "Sim scene: mean model-point error 0.019 mm with ground-truth poses and 0.20 cm with ICP poses
- ICP camera position error 0.87 cm and viewing-direction error 0.1 degrees on average (Sec. 7, Fig. 3).", "Large Office (about 10 m x 6 m x 2.5 m): 4.6 million model points stored in 110 MB of GPU memory, whereas a predefined 512 MB voxel grid would force voxels larger than 1 cm (Sec. 7).", "Tracking on Flowerpot is similar to KinectFusion against Vicon ground truth, with the largest difference about 1 cm (Fig. 5 caption).", "Dynamic segmentation keeps tracking in the Moving Person scene where earlier approaches fail, and model size converges after one turntable revolution (Sec. 7, Figs. 6 and 8)."]
限制
- ["Sensor drift is not tackled
- drift in larger environments remains future work (Sec. 8).", "Opaque splats are rendered without blending or prefiltering, trading local surface quality for speed (Sec. 5).", "Only the last RGB sample is stored per point (Sec. 7).", "Only the synthetic Sim scene has geometric ground truth
- the comparison with KinectFusion on Flowerpot and Teapot is visual apart from Vicon-based tracking (Sec. 7, Figs. 4-5).", "Hashing authors later state that point-based fusion quality is not on par with true volumetric methods (Niessner et al. 2013, Sec. 2)
- this is a competing-method claim, not an independent evaluation."]
營建工程相關證據
原研究為室內物件與辦公室尺度 RGB-D、ToF 場景（Large Office 約 10 m x 6 m x 2.5 m），未涉及營建。
原文驗證環境：模擬、受控實驗、獨立參考量測
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 40 筆紀錄。
Keller et al., 2013 · Table 1 本方法 35 筆
表格設定（擷取紀錄原文）：Average per-frame timings of ICP, dynamic segmentation and fusion; frames input/processed and fps input/processed; input 640x480 except PMD 200x200. Input fps: Sim 15, PMD 27, others 30 (Keller et al., 2013, Table 1)
Avg. timings [ms], ICP，Sim (synthetic) · 950/950 frames
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 Keller 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 |
|---|---|---|
| point-based fusion本方法原文提出硬體：Intel i7 8-core CPU, NVIDIA GTX 680 | 18.9 ms | (Keller et al., 2013, Table 1) |
Keller et al., 2013 · Text Sec. 7 本方法 5 筆
資料集與序列Sim (synthetic) · virtual camera rotating around the scene
表格設定（擷取紀錄原文）：Synthetic Sim scene with ground-truth camera transformations and geometry; errors are means over model points or frames (Keller et al., 2013, Text Sec. 7)
mean position error of global model points (ground-truth poses)，Sim (synthetic) · virtual camera rotating around the scene
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 Keller 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 |
|---|---|---|
| point-based fusion with ground-truth camera poses本方法原文提出 | 0.019 mm | (Keller et al., 2013, Sec. 7, Fig. 3) |
來源
Keller et al., 2013
（2013）Real-Time 3D Reconstruction in Dynamic Scenes Using Point-Based Fusion2013 International Conference on 3D Vision (3DV), pp. 1-8
同儕審查已出版已讀全文經典查證後修正