GS-LIVM
GS-LIVM 以改良的 SR-LIVO(ESIKF 緊耦合 LiDAR、慣性與視覺里程計)提供位姿,在體素層級以高斯過程回歸(Voxel-GPR)把稀疏且分布不均的 LiDAR 點轉為均勻網格點,並以預測變異數加權計算三維高斯的初始位置與尺度(旋轉設為單位四元數),再以影像、深度差與結構相似損失持續最佳化,在 8 GB 筆電 GPU 上完成大型戶外場景的即時寫實建圖。論文只評估渲染品質與軌跡,作者指出高斯初始化完全依賴點雲,LiDAR 未覆蓋處會出現缺漏。
本頁內容
Uses voxel-level GPR on sparse LiDAR points to initialize covariance-aware 3D Gaussians for real-time outdoor photorealistic mapping on top of LIVO poses.
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 | 3D LiDAR、IMU、monocular camera |
|---|---|
| 原文測試平台 | handheld、wheeled robot |
| 狀態估計 | ESIKF LiDAR-inertial-visual odometry (tracking thread, output at IMU rate) adopted from SR-LIVO [48] with two changes: original sensor timestamps replace the time-sweep refinement (fixing crashes with spinning LiDAR) and large matrix products are CUDA-accelerated |
| 資料關聯 | colour point cloud from LIVO; photometric, SSIM, structure-similarity and delta-depth losses for Gaussian optimization |
| 時間表示 | discrete poses |
| 去畸變 | 不適用 |
| 迴圈閉合 | none; the authors state the tracking method lacks a loop closure detection module and is less accurate than LVI-SAM on some Botanic Garden sequences |
| 全域最佳化 | none reported |
| 地圖表示 | voxel-hashed dense map of 3D Gaussians (position, covariance, opacity, zero-degree SH colour) initialized per voxel subgrid from Voxel-GPR predictions: position as the inverse-variance weighted mean, scale from the diagonal of the weighted covariance, rotation set to the identity quaternion |
| 先驗資訊 | none |
| 可輸出幾何 | dense 3D Gaussian map and rendered images; no geometric accuracy metric of the map is reported (evaluation covers PSNR, SSIM, LPIPS, runtime, memory and trajectory RPE and ATE only) |
| 計算需求 | CUDA/C++ with LibTorch under ROS; desktop PC with a 5.50 GHz Intel Core i9-13900HX CPU, 64 GB RAM and an NVIDIA RTX 4060 Laptop 8 GB GPU; Voxel-GPR under 30 ms per batch; all sequences fully mapped on the 8 GB GPU with ns = 3; mapping time equals or nearly equals the sequence duration (e.g., 612 s for a 611 s sequence) |
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 |
|---|---|---|---|---|---|
| LiDAR | Livox Avia | 資料集感測器 | R3LIVE dataset; FAST-LIVO dataset | 10 Hz; handheld device with built-in IMU and a 15 Hz RGB camera (640 x 512) | (Xie et al., 2025, Sec. 4.1) |
| LiDAR | Ouster-16 | 資料集感測器 | NTU-VIRAL | multi-line spinning LiDAR, sparser than Livox; images 752 x 480 | (Xie et al., 2025, Sec. 4.1) |
| LiDAR | Velodyne VLP-16 | 資料集感測器 | Botanic Garden | multi-line spinning LiDAR; images 480 x 300; used for the tracking evaluation | (Xie et al., 2025, Sec. 4.1; Supp. 7.4) |
| LiDAR | Livox Avia | 資料集感測器 | Botanic Garden | second LiDAR on the Botanic Garden robot | (Xie et al., 2025, Sec. 4.1) |
| LiDAR | Livox MID-360 | 方法輸入 | 未標示 | self-collected sequence 'private-360', images 640 x 512 | (Xie et al., 2025, Supp. 7.2) |
| LiDAR | HESAI Pandar XT-32歸入:Hesai PandarXT-32 | 方法輸入 | 未標示 | self-collected sequence 'private-pandar', images 640 x 512 | (Xie et al., 2025, Supp. 7.2) |
| 慣性量測單元(IMU) | Livox Avia built-in IMU | 資料集感測器 | R3LIVE dataset; FAST-LIVO dataset | 200 Hz | (Xie et al., 2025, Sec. 4.1) |
| 相機 | RGB camera (model not reported) | 資料集感測器 | R3LIVE dataset; FAST-LIVO dataset | 15 Hz, 640 x 512 | (Xie et al., 2025, Sec. 4.1) |
| 相機 | camera (model not reported) | 資料集感測器 | NTU-VIRAL | image resolution 752 x 480 | (Xie et al., 2025, Sec. 4.1) |
| 相機 | camera (model not reported) | 資料集感測器 | Botanic Garden | image resolution 480 x 300 | (Xie et al., 2025, Sec. 4.1) |
| 載具平台 | wheeled robot | 資料集感測器 | Botanic Garden | traverses a botanic garden | (Xie et al., 2025, Sec. 4.1) |
| 載具平台 | handheld device | 資料集感測器 | R3LIVE dataset; FAST-LIVO dataset | Livox Avia, built-in IMU and RGB camera | (Xie et al., 2025, Sec. 4.1) |
| 運算硬體 | Intel Core i9-13900HX | 執行運算平台 | 未標示 | 5.50 GHz CPU, 64 GB RAM desktop PC | (Xie et al., 2025, Sec. 4.1 Implementation Details) |
| 運算硬體 | NVIDIA RTX 4060 Laptop 8 GB GPU | 執行運算平台 | 未標示 | 8 GB | (Xie et al., 2025, Sec. 4.1 Implementation Details; Sec. 4.3) |
作者報告的優勢與限制
優勢
- Real-time photorealistic mapping in large unbounded outdoor scenes (abstract)
- Best LPIPS on all four Table 1 sequences and best PSNR on three of four under the 100-frame, 5-minute comparison protocol (Table 1)
- Full sequences mapped on an 8 GB laptop GPU, including a 1170 s, about 900 m sequence in real time with nr = 2 (Sec. 4.3; Supp. 7.2, 7.5; Table 8)
- The authors' improved SR-LIVO tracker has the lowest RPE on all seven Botanic Garden sequences and the lowest ATE on two of seven (Supp. Table 7)
限制
- Initialization relies solely on point clouds, leaving missing regions without LiDAR coverage (Supp. 8)
- Real-time constraints limit reconstruction quality (Supp. 8)
- No loop closure; ATE higher than LVI-SAM on five of seven Botanic Garden sequences (Supp. 7.4, Table 7)
- A fully optimized offline 3DGS scores higher on training views (hku seq 00 PSNR 27.827 vs 22.430), although the authors report poorer novel-view synthesis for it (Supp. 7.1, Table 4, Fig. 8)
- Rendering from sparse spinning LiDAR is worse than from Livox LiDAR; 3DGS has higher PSNR than GS-LIVM on all five additional sequences of Supp. Table 5 (Supp. 7.2)
- Baselines in Table 1 were run on reduced scenes (at most 100 frames, 5 min) with COLMAP poses and depth, while GS-LIVM used its own odometry (Sec. 4.2)
- Sky regions receive no LiDAR returns and are excluded; handling them is left to future work (ICCV Supp. 8; absent from arXiv v1)
營建工程相關證據
論文未在營建場域測試;資料為 R3LIVE、FAST-LIVO、NTU-VIRAL、Botanic Garden,以及作者以 Livox MID-360 與 HESAI Pandar XT-32 自錄的兩段序列。論文只評估渲染品質、資源用量與軌跡,未量測地圖幾何精度,因此無法直接支持營建量測用途。
原文驗證環境:公開基準
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 9 個比較組,合計 64 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 5 組列在最後,並連到性能比較頁。
Xie et al., 2025 · Supp. Table 7 本方法 14 筆
表格設定(擷取紀錄原文):Tracking accuracy on Botanic Garden using Velodyne VLP-16 data; RPE and ATE over full transformations; units, statistic and alignment not stated; 'Ours' is the authors' improved SR-LIVO tracker without loop closure; arXiv v1 supplementary (Xie et al., 2025, Supp. Table 7)
RPE (full transformation),Botanic Garden · 1005 00
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 Xie et al., 2025 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(Xie et al., 2025, Supp. Table 7)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| R3LIVE [18] | 1.165 | (Xie et al., 2025, arXiv v1 Supp. Table 7) |
| FAST-LIO2 [43] | 1.048 | (Xie et al., 2025, arXiv v1 Supp. Table 7) |
| LVI-SAM [33] | 0.347 | (Xie et al., 2025, arXiv v1 Supp. Table 7) |
| OURS本方法原文提出 | 0.174 | (Xie et al., 2025, arXiv v1 Supp. Table 7) |
Xie et al., 2025 · Table 1 本方法 12 筆
表格設定(擷取紀錄原文):Rendering quality, mean over all observation images; reduced scenes of at most 100 frames and 5 min reconstruction; NeRF-SLAM, MonoGS and 3DGS given COLMAP poses and depth, GS-LIVM uses its own LIVO poses (Xie et al., 2025, Table 1)
PSNR,R3LIVE dataset · hku campus seq 00
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 Xie et al., 2025 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(Xie et al., 2025, Table 1)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Nerf-SLAM [27] | 13.232 dB | (Xie et al., 2025, Table 1) |
| MonoGS [21] | 12.142 dB | (Xie et al., 2025, Table 1) |
| 3DGS [13] | 21.744 dB | (Xie et al., 2025, Table 1) |
| Ours本方法原文提出 | 22.43 dB | (Xie et al., 2025, Table 1) |
Xie et al., 2025 · Table 2 本方法 12 筆
表格設定(擷取紀錄原文):Mapping time (MT) against sequence duration (DT), number of 3D Gaussians and maximum GPU memory for the full sequence, ns = 3 (Xie et al., 2025, Table 2)
Count of 3D Gaussians,R3LIVE dataset · hku campus seq 00
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 Xie et al., 2025 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 |
|---|---|---|
| Ours本方法原文提出硬體:Intel Core i9-13900HX (5.50 GHz), 64 GB RAM, NVIDIA RTX 4060 Laptop 8 GB GPU | 1209666 count | (Xie et al., 2025, Table 2) |
Xie et al., 2025 · Supp. Table 8 本方法 8 筆
表格設定(擷取紀錄原文):Ultra-long sequences with nr = 2: mapping time vs duration, peak memory and rendering metrics; arXiv v1 supplementary (Xie et al., 2025, Supp. Table 8)
Mem (Mb),R3LIVE or FAST-LIVO dataset (not stated) · hku main building
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 Xie et al., 2025 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 |
|---|---|---|
| Ours本方法原文提出硬體:Intel Core i9-13900HX (5.50 GHz), 64 GB RAM, NVIDIA RTX 4060 Laptop 8 GB GPU | 4489 | (Xie et al., 2025, arXiv v1 Supp. Table 8) |
其他比較組
來源
Xie et al., 2025
(2025)GS-LIVM: Real-Time Photo-Realistic LiDAR-Inertial-Visual Mapping with Gaussian Splatting2025 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 26869-26878
DOI 10.1109/iccv51701.2025.02494arXiv 2410.17084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:2410.17084 https://arxiv.org/abs/2410.17084
程式碼:https://github.com/xieyuser/GS-LIVM(授權:GPL-3.0)。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。