LiV-GS
LiV-GS 以點雲與高斯共有的共變異為橋樑,直接把稀疏 LiDAR 點與連續可微的高斯地圖對齊做前端追蹤,並對 LiDAR 視野外的高斯施加條件約束,使其貼近鄰近可靠高斯。軌跡評估以 R3LIVE 的軌跡作為參考真值,幾何精度只以毫米波雷達跨模態重定位做定性說明。
本頁內容
Aligns LiDAR scans directly to a 3D Gaussian map through covariance for tracking, with conditional constraints for Gaussians outside LiDAR coverage.
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 (Livox Horizon in NTU4DRadLM; Livox Avia in R3LIVE hku_park_00)、monocular camera (640 x 480 in NTU4DRadLM; 1280 x 1024 at 30 Hz in hku_park_00) |
|---|---|
| 原文測試平台 | handheld、vehicle |
| 狀態估計 | frame-to-map alignment of LiDAR points to Gaussians through shared covariance/normal attributes; sliding-window back-end pose and map optimization |
| 資料關聯 | point-to-Gaussian (covariance-based plane) matching |
| 時間表示 | discrete poses |
| 去畸變 | 不適用 |
| 迴圈閉合 | none (authors state LiV-GS lacks loop closure detection) |
| 全域最佳化 | none |
| 地圖表示 | 3D Gaussians with normal-aware loss and conditional Gaussian constraints outside the LiDAR field of view |
| 先驗資訊 | none (IMU not used) |
| 可輸出幾何 | Gaussian map and renderings; geometry accuracy assessed only qualitatively via cross-modal radar relocalization |
| 計算需求 | 7.98 FPS overall on a desktop with an NVIDIA RTX 4090 GPU; front-end tracking and keyframe selection 0.07 ms, pose optimization 0.04 ms (five keyframes) and map update 0.09 ms per operation, with asynchronous modules; IMU not used |
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-Horizon (as written)歸入:Livox Horizon | 資料集感測器 | NTU4DRadLM | 10 Hz | (Xiao et al., 2025, Sec. IV-B) |
| LiDAR | Livox-Avia (as written)歸入:Livox Avia | 資料集感測器 | R3LIVE dataset (hku_park_00) | 10 Hz; handheld scanning | (Xiao et al., 2025, Sec. IV-G) |
| 相機 | monocular camera (model not reported) | 資料集感測器 | NTU4DRadLM | 640 x 480 | (Xiao et al., 2025, Sec. IV-B) |
| 相機 | camera of the R3LIVE handheld device (model not reported) | 資料集感測器 | R3LIVE dataset (hku_park_00) | 1280 x 1024 at 30 Hz | (Xiao et al., 2025, Sec. IV-G) |
| 雷達 | Eagle Ocuill G7 (as written)歸入:Eagle Ocuill G7 | 資料集感測器 | NTU4DRadLM | 4D millimeter-wave radar; used only for cross-modal relocalization on the Gaussian map | (Xiao et al., 2025, Sec. IV-B; Sec. IV-F) |
| 載具平台 | human-driven vehicle platform | 資料集感測器 | NTU4DRadLM | used for the high-speed loop2 sequence | (Xiao et al., 2025, Sec. IV-B) |
| 運算硬體 | NVIDIA RTX 4090 GPU | 執行運算平台 | 未標示 | desktop; all compared algorithms run on it | (Xiao et al., 2025, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Lowest ATE on all five low-speed NTU4DRadLM segments (0.366 m to 0.771 m) (Sec. IV-C, Table II)
- With its own odometry, higher rendering SSIM than every other compared method on all six NTU4DRadLM sequences; its ground-truth-pose variant scores higher on nyl2 (0.744 vs 0.725) (Table III)
- Lowest t_rel and r_rel on the closed-loop hku_park_00 segment (Table IV)
限制
- No loop closure; on the closed-loop hku_park_00 segment ATE is 0.705 m vs 0.457 m for ORB-SLAM3 (Sec. IV-G, Table IV)
- On high-speed loop2 the authors report accuracy slightly below NeRF-LOAM, but Table II shows HDL-graph-SLAM with the lowest ATE (0.593 m vs 0.843 m) and NeRF-LOAM better only in rotational drift (Sec. IV-C, Table II)
- No spherical harmonics, so view-dependent colour at loop closures can create overlapping Gaussians; on hku_park_00 rendering PSNR is 17.316 vs 25.776 for 3DGS (Sec. IV-G, Table V)
- Trajectory reference is R3LIVE output rather than an independent measurement (Sec. IV-A) (inference: reference uncertainty not quantified)
- Geometry accuracy shown only qualitatively through mmWave radar relocalization with HDL-localization (Sec. IV-F)
營建工程相關證據
論文未在營建場域測試;資料為 NTU4DRadLM 與 R3LIVE hku park 序列。
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 4 個比較組,合計 29 筆紀錄。
Xiao et al., 2025 · Table II 本方法 18 筆
表格設定(擷取紀錄原文):Tracking accuracy with rpg trajectory evaluation: t_rel = average translational RMSE drift (%), r_rel = average rotational RMSE drift (deg/100 m), t_abs = ATE RMSE (m); reference trajectories from R3LIVE (not an independent measurement); alignment not stated; IMU not used by LiV-GS; '-' entries reported without explanation (text says indoor-oriented 3DGS SLAM methods degrade or fail on some outdoor sequences) (Xiao et al., 2025, Table II)
t_rel (average translational RMSE drift),NTU4DRadLM · cp
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。
- 未報告(沒有數值,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 Xiao 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(Xiao et al., 2025, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| NeRF-LOAM | 2.943% | (Xiao et al., 2025, Table II) |
| HDL-graph-SLAM | 1.264% | (Xiao et al., 2025, Table II) |
| ORB-SLAM3 | 1.356% | (Xiao et al., 2025, Table II) |
| SplaTAM | 無數值未報告註記(擷取紀錄):not reported ('-' in table) | (Xiao et al., 2025, Table II) |
| MonoGS | 4.171% | (Xiao et al., 2025, Table II) |
| Gaussian-SLAM | 1.249% | (Xiao et al., 2025, Table II) |
| GS-ICP-SLAM | 5.471% | (Xiao et al., 2025, Table II) |
| Ours本方法原文提出 | 0.234% | (Xiao et al., 2025, Table II) |
Xie et al., 2025 · ICCV Supp. Table 4 本方法 4 筆
表格設定(擷取紀錄原文):Rendering comparison with LiV-GS on sequences cp and nyl2; LiV-GS values copied from its preprint (marked *), GS-LIVM values from the authors' runs; dataset not named in GS-LIVM (the sequence names match NTU4DRadLM segments in LiV-GS); added by second checker from the ICCV supplement (Xie et al., 2025, ICCV Supp. Table 4)
PSNR,not stated · cp
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, ICCV Supp. Table 4)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| LiV-GS* (preprint result)本方法 | 22.27 dB | (Xie et al., 2025, ICCV Supp. Table 4) |
| Ours原文提出 | 23.54 dB | (Xie et al., 2025, ICCV Supp. Table 4) |
Xiao et al., 2025 · Text Sec. IV-E 本方法 4 筆
資料集與序列not stated (Fig. 7 relates runtime to the NTU4DRadLM accuracy and rendering results, but the runtime data are not attributed to a dataset) · not stated
表格設定(擷取紀錄原文):System throughput (processed frames divided by total time) and mean per-operation module runtimes; modules run asynchronously (Xiao et al., 2025, Text Sec. IV-E)
FPS (depth maps and RGB images processed per second),not stated (Fig. 7 relates runtime to the NTU4DRadLM accuracy and rendering results, but the runtime data are not attributed to a dataset) · not stated
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 Xiao 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:desktop with NVIDIA RTX 4090 GPU | 7.98 fps | (Xiao et al., 2025, Abstract; Sec. IV-E; Fig. 7) |
Xiao et al., 2025 · Table IV 本方法 3 筆
資料集與序列R3LIVE dataset · hku_park_00 (closed-loop segment)
表格設定(擷取紀錄原文):Closed-loop test on a segment of R3LIVE hku_park_00 (handheld, Livox Avia 10 Hz, 1280 x 1024 images at 30 Hz); same metrics as Table II; LiV-GS has no loop closure detection (Xiao et al., 2025, Table IV)
t_rel (average translational RMSE drift),R3LIVE dataset · hku_park_00 (closed-loop segment)
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。
- 未報告(沒有數值,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 Xiao 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(Xiao et al., 2025, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| NeRF-LOAM | 0.518% | (Xiao et al., 2025, Table IV) |
| HDL-graph-SLAM | 0.497% | (Xiao et al., 2025, Table IV) |
| ORB-SLAM3 | 0.446% | (Xiao et al., 2025, Table IV) |
| SplaTAM | 無數值未報告註記(擷取紀錄):not reported ('-' in table) | (Xiao et al., 2025, Table IV) |
| MonoGS | 0.535% | (Xiao et al., 2025, Table IV) |
| Gaussian-SLAM | 1.456% | (Xiao et al., 2025, Table IV) |
| GS-ICP-SLAM | 無數值未報告註記(擷取紀錄):not reported ('-' in table) | (Xiao et al., 2025, Table IV) |
| Ours本方法原文提出 | 0.436% | (Xiao et al., 2025, Table IV) |
來源
Xiao et al., 2025
(2025)LiV-GS: LiDAR-Vision Integration for 3D Gaussian Splatting SLAM in Outdoor EnvironmentsIEEE Robotics and Automation Letters, 10(1), 421-428
DOI 10.1109/lra.2024.3505777arXiv 2411.12185
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:2411.12185 https://arxiv.org/abs/2411.12185