Vox-Fusion
Vox-Fusion 將神經隱式表面與傳統體素融合結合：場景以八元樹（octree）與 Morton 編碼管理的稀疏體素表示，體素頂點存放共享的特徵向量，再由多層感知器解碼成 SDF 與顏色。新影格的深度點雲一旦落在既有體素之外就即時配置新體素，因此不必預先知道場景範圍，記憶體也只花在有觀測的表面附近。追蹤時固定地圖，只以可微分體積渲染最佳化相機位姿；建圖時則對隨機挑選的關鍵影格視窗聯合最佳化地圖與位姿，但系統沒有迴圈閉合或全域最佳化。
本頁內容
RGB-D neural implicit SLAM whose SDF map is stored as embeddings on a sparse, dynamically allocated octree of voxels, so it needs no predefined scene bound; tracking and windowed joint mapping both use differentiable SDF volume rendering, without loop closure.
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 (synthetic Replica, ScanNet, iPhone 13 Pro and iPad Pro (2020) with onboard LiDAR depth) |
|---|---|
| 原文測試平台 | simulation (Replica synthetic RGB-D sequences)、handheld (iPhone 13 Pro and iPad Pro captures, Fig. 8)、原文未報告 (ScanNet capture platform not described) |
| 狀態估計 | gradient-based 6-DoF pose optimization in se(3) through differentiable SDF volume rendering against a frozen copy of the map (zero-motion initialization); mapping jointly optimizes decoder, voxel embeddings and poses of a random keyframe window (Sec. 4.2-4.3) |
| 資料關聯 | direct: rendered colour and depth losses plus free-space and SDF losses on sparsely sampled pixels whose rays hit allocated voxels (Sec. 4.1) |
| 時間表示 | discrete poses |
| 去畸變 | 不適用 |
| 迴圈閉合 | none (the authors list drift in long-time tracking as unresolved, Sec. 7) |
| 全域最佳化 | none; joint optimization only over a window of randomly selected keyframes (Sec. 4.3) |
| 地圖表示 | sparse voxel grid (voxel size 0.2 m) in an octree with Morton coding, 16-D embeddings on voxel vertices shared by neighbours, decoded by an MLP into SDF and colour; voxels allocated on the fly from back-projected depth, so no scene bound is needed (Sec. 4.4, Sec. 5.1) |
| 先驗資訊 | none; decoder and embeddings are learned on the fly without pre-trained geometry priors (Sec. 2.1) |
| 可輸出幾何 | camera trajectory and SDF-based surface mesh (evaluated as mesh against ground-truth mesh); rendered colour and depth images (Sec. 5) |
| 計算需求 | single NVIDIA RTX 3090; about 150-200 ms to track a frame and 450-550 ms per joint optimization, i.e. about 5 Hz tracking and 2 Hz mapping (Sec. 5.4, Table 4) |
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 | iPhone 13 Pro | 方法輸入 | 未標示 | RGB images with depth from the onboard lidar sensor; very low depth resolution | (Yang et al., 2022, Sec. 5.1, Sec. 5.3, Fig. 8) |
| RGB-D 相機 | iPad Pro (2020) | 方法輸入 | 未標示 | iOS device with range sensor used for RGB-D capture | (Yang et al., 2022, Sec. 5.1, Sec. 5.3) |
| 運算硬體 | NVidia RTX 3090歸入：NVIDIA RTX 3090 | 執行運算平台 | 未標示 | single video card used for profiling | (Yang et al., 2022, Sec. 5.4) |
作者報告的優勢與限制
優勢
- Dynamic sparse voxel allocation removes the need for a known scene bound, unlike NICE-SLAM's pre-allocated dense grid (Sec. 2.1, Sec. 4.4)
- Embedding memory on Replica office-0 is 0.149 MB versus 238.88 MB for NICE-SLAM (Table 5)
- Replica average ATE RMSE 0.0054 m versus 0.0195 m for NICE-SLAM, and average accuracy 2.37 cm, completion 2.28 cm, completion ratio 92.86% (Tables 1-2)
- Leaves large unobserved regions empty instead of hallucinating surfaces, so observed and unobserved space remain distinguishable (Sec. 5.2)
限制
- Cannot robustly handle dynamic objects or drift in long-time tracking (Sec. 7)
- No loop closure or global optimization (Sec. 4.3, Sec. 7) (inference from the method description)
- Typical rates are about 5 Hz tracking and 2 Hz mapping on an RTX 3090, and may be slower in challenging scenes (Sec. 5.4)
- Low-resolution iOS depth limits reconstruction quality of the handheld outdoor and indoor captures, which are shown only qualitatively (Sec. 5.3, Fig. 8)
營建工程相關證據
論文未涉及營建場域；定量評估只在 Replica 合成場景與 ScanNet 五個室內場景進行，iPhone 13 Pro 與 iPad Pro 的室外手持掃描僅作定性展示。其稀疏體素依觀測動態擴張、不需預設場景邊界的設計，較適合範圍事先未知的工地（推論），但系統無迴圈閉合，作者也承認長時間追蹤會漂移，大範圍施工現場仍需額外的全域校正。
原文驗證環境：公開基準、模擬
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
Yang et al., 2022
（2022）Vox-Fusion: Dense Tracking and Mapping with Voxel-based Neural Implicit Representation2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pp. 499-507
DOI 10.1109/ismar55827.2022.00066arXiv 2210.15858程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本：arXiv:2210.15858v3 https://arxiv.org/abs/2210.15858
- 程式碼釋出：zju3dv/Vox-Fusion https://github.com/zju3dv/Vox-Fusion
程式碼：https://github.com/zju3dv/Vox-Fusion（授權：原文未報告 (no LICENSE file in the repository; GitHub reports no 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。