HI-SLAM2
HI-SLAM2 是只用單眼 RGB 的三維高斯 SLAM：追蹤端沿用 DROID-SLAM 的學習式光流與稠密光束法平差，並以每張影像 2x2 的尺度網格把 Omnidata 單眼深度先驗對齊到估計深度，以修正先驗中隨位置變化的尺度失真。偵測到迴圈時以 Sim(3) 位姿圖平差同時修正位姿與尺度漂移，並依錨定關鍵影格的更新直接變形高斯，使地圖即時保持一致；離線階段再做完整光束法平差與位姿、高斯聯合最佳化。地圖以高斯表示並以射線與高斯交點計算無偏深度，最後由渲染深度經 TSDF 融合得到網格。
本頁內容
Monocular RGB Gaussian-splatting SLAM that couples DROID-SLAM tracking with grid-based scale alignment of monocular depth priors, Sim(3) pose-graph loop closure that deforms keyframe-anchored Gaussians online, and offline full BA plus joint pose-map refinement; meshes via TSDF fusion of rendered depth.
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 | monocular RGB camera |
|---|---|
| 原文測試平台 | simulation (Replica)、vehicle (Waymo Open front camera)、ground robot with stereo cameras, left camera only (self-collected factory hall; robot type not described)、handheld (ScanNet++ iPhone sequences) |
| 狀態估計 | DROID-SLAM-based recurrent optical flow and dense bundle adjustment (damped Gauss-Newton) on a keyframe graph, interleaved with joint depth and scale alignment (JDSA) that fits a 2x2 scale grid to each monocular depth prior; Sim(3) pose-graph BA on loop closure; offline full BA and joint pose and 3DGS refinement with Adam (Sec. III-B to III-E) |
| 資料關聯 | dense learned optical-flow correspondences with confidence weights between co-visible keyframes (Sec. III-B) |
| 時間表示 | discrete poses (keyframes) |
| 去畸變 | 不適用 |
| 迴圈閉合 | proximity-based: candidates with optical-flow distance below a threshold, orientation difference below a threshold and index gap beyond the local window; closed by online Sim(3) pose-graph bundle adjustment that also corrects scale drift (Sec. III-C) |
| 全域最佳化 | online Sim(3) pose-graph BA with relative-pose factors from retained dense correspondences, offline full BA over all overlapping keyframe pairs, then joint optimization of Gaussians, poses and exposure (Sec. III-C, III-E) |
| 地圖表示 | 3D Gaussian splatting with RGB colours (no spherical harmonics), unbiased ray-Gaussian intersection depth, Gaussians anchored to keyframes and deformed with their Sim(3) updates; random downsampling (factor 32), densification and pruning; map grows without a predefined scene bound (Sec. III-D) |
| 先驗資訊 | pretrained DROID-SLAM flow network and Omnidata monocular depth and normal priors (Sec. IV-A) |
| 可輸出幾何 | 3DGS map and a mesh from TSDF fusion of rendered depth maps; rendered colour and depth (Sec. III) |
| 計算需求 | NVIDIA RTX 4090 and Intel Core i9-12900K; online tracking, loop closing and mapping at 22 FPS on Replica and above 10 FPS on ScanNet; offline refinement 12 s on Replica and a few minutes on ScanNet (Sec. IV-A, IV-H) |
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 | Nvidia RTX 4090歸入：NVIDIA RTX 4090 | 執行運算平台 | 未標示 | GPU for all evaluations | (Zhang et al., 2025, Sec. IV-A) |
| 運算硬體 | Intel Core i9-12900K歸入：Intel Core i9 12900K | 執行運算平台 | 未標示 | CPU for all evaluations | (Zhang et al., 2025, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Best Replica geometry among RGB-only methods: accuracy 1.57 cm, completeness 3.49 cm, completion ratio 85.25% (Table IV)
- Lowest average ATE on Replica (0.26 cm), ScanNet (7.07 cm) and Waymo Open (0.457 m) among compared methods (Tables I-III)
- Online operation at 22 FPS on Replica with 12 s offline refinement (Sec. IV-H)
- Factory-hall robot sequence processed in about half the runtime of DROID-SLAM + 3DGS with better geometry; trajectory compared against a photogrammetric reference (Sec. IV-I, Fig. 14)
限制
- Proximity-based loop detection is not robust to view occlusions and textureless regions (ETH3D); learned place recognition is suggested (Sec. V)
- Mapping quality can degrade in city-scale scenes because of the limited optimization budget; submaps suggested (Sec. V)
- Assumes static environments (Sec. V)
- On ETH3D, 6 of 61 sequences failed in complete darkness and 4 more because of lighting changes and occlusions (Sec. IV-D)
營建工程相關證據
論文在大型工廠廠房以機器人左相機拍攝 4073 影格進行實測，並以文獻 [81] 的公分級攝影測量參考比對軌跡（僅圖 14 定性呈現），也在具雷射掃描真值的 ScanNet++ iPhone 序列上展示重建；作者來自斯圖加特大學攝影測量研究所。其前身 HI-SLAM 是語料中 AEC 領域 SLAM 回顧（Li et al., 2026a）唯一提及的神經 SLAM。定量幾何精度仍只在合成 Replica 上報告（精度約 1.6 cm），在施工現場或以全測站、TLS 驗證的量測尚未出現；作者也指出城市尺度與動態環境仍是限制。
原文驗證環境：公開基準、模擬、受控實驗
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
Zhang et al., 2025
（2025）HI-SLAM2: Geometry-Aware Gaussian SLAM for Fast Monocular Scene ReconstructionIEEE Transactions on Robotics, 41:6478-6493
DOI 10.1109/tro.2025.3626627arXiv 2411.17982程式碼
同儕審查已出版已讀全文近十年
相關版本
- 預印本：arXiv:2411.17982v3 https://arxiv.org/abs/2411.17982
- predecessor：HI-SLAM (IEEE RA-L 2024) https://doi.org/10.1109/LRA.2023.3347131
- 程式碼釋出：Willyzw/HI-SLAM2 https://github.com/Willyzw/HI-SLAM2
程式碼：https://github.com/Willyzw/HI-SLAM2（授權：BSD-3-Clause (LICENSE file checked; copyright line inherited from Princeton Vision & Learning Lab)）。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。