Gaussian-LIC
Gaussian-LIC 以連續時間緊耦合的 LiDAR、慣性與相機里程計(Coco-LIC,每 0.1 秒做一次因子圖最佳化)提供位姿,將著色並降取樣的 LiDAR 點與視覺滑動視窗三角化的 SfM 點一起初始化三維高斯,以補足 LiDAR 未涵蓋的相機視野,並加入天空高斯與曝光仿射模型,以 C++ 與 CUDA 加速達成即時寫實建圖。論文只報告渲染品質(PSNR、SSIM、LPIPS)與執行時間,追蹤比較僅為定性描述,沒有 ATE 或地圖幾何精度數值;作者把提升幾何重建品質列為未來工作。
本頁內容
Tightly coupled continuous-time LiDAR-inertial-camera odometry feeds LiDAR and triangulated points into an online 3DGS map for real-time photorealistic mapping.
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 | 未記錄 |
| 狀態估計 | continuous-time factor-graph sliding-window optimization (Coco-LIC) with point-to-map LiDAR, frame-to-map visual and inertial factors |
| 資料關聯 | Coco-LIC odometry with point-to-map LiDAR factors, frame-to-map visual factors and inertial factors (Sec. III-B); a separate VINS-Mono-style visual sliding window tracks Shi-Tomasi corners with KLT only to triangulate SfM points for Gaussian initialization; mapping minimizes an L1 plus D-SSIM re-rendering loss with a per-image exposure affine matrix (Eq. 9) |
| 時間表示 | continuous-time (Coco-LIC) |
| 去畸變 | not described in the full text (v3); poses come from the continuous-time Coco-LIC trajectory optimized every 0.1 s |
| 迴圈閉合 | none reported |
| 全域最佳化 | no pose graph, global BA or loop closure; the Gaussian map is optimized at each keyframe on K = 100 keyframes sampled from all keyframes to limit forgetting |
| 地圖表示 | 3D Gaussians initialized from colourized LiDAR points plus triangulated visual SfM points; sky and exposure modelling |
| 先驗資訊 | none |
| 可輸出幾何 | 3D Gaussian map (with sky Gaussians) and rendered images; neither map geometric accuracy nor trajectory accuracy is quantified; authors list improving geometric reconstruction quality as future work |
| 計算需求 | desktop with NVIDIA RTX 3090 (24 GB), Intel Core i7-8700 (3.2 GHz) and 32 GB RAM; C++/CUDA with LibTorch and ROS; on sequence f0 (105 s) tracking and mapping both finish in 105 s (198 s without the acceleration strategies), the only real-time method compared |
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 | solid-state LiDAR (model not named) | 資料集感測器 | FAST-LIVO dataset | used for Gaussian initialization and LiDAR factors | (Lang et al., 2025, Sec. IV-A2) |
| LiDAR | solid-state LiDAR (model not named) | 資料集感測器 | R3LIVE dataset | used for Gaussian initialization and LiDAR factors | (Lang et al., 2025, Sec. IV-A2) |
| LiDAR | mechanical spinning LiDAR (model not named) | 資料集感測器 | MCD dataset | large-scale MCD dataset; segments of tuhh_day_02, tuhh_day_03 and tuhh_day_04 | (Lang et al., 2025, Sec. IV-A2) |
| 慣性量測單元(IMU) | IMU (model not named) | 資料集感測器 | FAST-LIVO, R3LIVE and MCD datasets | inertial factors in Coco-LIC | (Lang et al., 2025, Sec. III-B; Sec. IV-A2) |
| 相機 | RGB camera (model not named) | 資料集感測器 | FAST-LIVO dataset; R3LIVE dataset | 640x512 images; only left images used when stereo is provided | (Lang et al., 2025, Sec. IV-A2) |
| 相機 | RGB camera (model not named) | 資料集感測器 | MCD dataset | 640x480 images; only left images used when stereo is provided | (Lang et al., 2025, Sec. IV-A2) |
| 運算硬體 | NVIDIA RTX 3090 | 執行運算平台 | 未標示 | 24 GB VRAM | (Lang et al., 2025, Sec. IV-A1) |
| 運算硬體 | Intel Core i7-8700 | 執行運算平台 | 未標示 | 3.2 GHz CPU, 32 GB RAM | (Lang et al., 2025, Sec. IV-A1) |
作者報告的優勢與限制
優勢
- Robust tracking where RGB-only and RGB-D 3DGS baselines drift or fail (Sec. IV-B1)
- Real-time processing within sensor duration (Sec. IV-B3)
限制
- No quantitative trajectory accuracy (ATE) or map geometric accuracy is reported; the tracking comparison is qualitative (Sec. IV-B1)
- Offline COLMAP plus 3DGS reaches higher PSNR on f0 (32.56 versus 29.89 dB) though not in real time (Table II)
- Authors plan to improve odometry with the Gaussian map and to improve geometric reconstruction quality (Sec. V)
營建工程相關證據
論文未在營建場域測試;資料為 FAST-LIVO、R3LIVE(固態 LiDAR)與 MCD(旋轉式 LiDAR)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 37 筆紀錄。
Lang et al., 2025 · Table I 本方法 24 筆
表格設定(擷取紀錄原文):Rendering quality; compared methods mapped with ground-truth poses (MCD) or Gaussian-LIC estimated poses (FAST-LIVO, R3LIVE); FAST-LIVO and R3LIVE use a solid-state LiDAR, MCD a spinning LiDAR (Lang et al., 2025, Table I)
PSNR (dB),FAST-LIVO · f0 hku2
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 Lang 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(Lang et al., 2025, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| NeRF-SLAM (train view) | 25.56 dB | (Lang et al., 2025, Table I) |
| MonoGS (train view) | 23.58 dB | (Lang et al., 2025, Table I) |
| SplaTAM with LiDAR pseudo RGB-D (train view) | 25.51 dB | (Lang et al., 2025, Table I) |
| Gaussian-LIC (train view)本方法原文提出 | 29.89 dB | (Lang et al., 2025, Table I) |
| Gaussian-LIC (novel view)本方法原文提出 | 29.28 dB | (Lang et al., 2025, Table I) |
Lang et al., 2025 · Table II 本方法 8 筆
資料集與序列FAST-LIVO · f0 hku2 (duration 105 s)
表格設定(擷取紀錄原文):Runtime on sequence f0 (105 s of data) with each method's own estimated poses; real time means finishing within the data duration (Lang et al., 2025, Table II)
Tracking time (s),FAST-LIVO · f0 hku2 (duration 105 s)
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。
- 不適用
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 Lang 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(Lang et al., 2025, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| NeRF-SLAM硬體:NVIDIA RTX 3090 (24 GB) + Intel Core i7-8700 (3.2 GHz), 32 GB RAM | 105 s | (Lang et al., 2025, Table II) |
| MonoGS硬體:NVIDIA RTX 3090 (24 GB) + Intel Core i7-8700 (3.2 GHz), 32 GB RAM | 181 s | (Lang et al., 2025, Table II) |
| SplaTAM硬體:NVIDIA RTX 3090 (24 GB) + Intel Core i7-8700 (3.2 GHz), 32 GB RAM | 954 s | (Lang et al., 2025, Table II) |
| COLMAP + 3DGS (offline)硬體:NVIDIA RTX 3090 (24 GB) + Intel Core i7-8700 (3.2 GHz), 32 GB RAM | 無數值不適用註記(擷取紀錄):不適用 (offline batch pipeline) | (Lang et al., 2025, Table II) |
| Gaussian-LIC本方法原文提出硬體:NVIDIA RTX 3090 (24 GB) + Intel Core i7-8700 (3.2 GHz), 32 GB RAM | 105 s | (Lang et al., 2025, Table II) |
| Gaussian-LIC w/o acceleration本方法原文提出硬體:NVIDIA RTX 3090 (24 GB) + Intel Core i7-8700 (3.2 GHz), 32 GB RAM | 105 s | (Lang et al., 2025, Table II) |
Lang et al., 2025 · Table III 本方法 4 筆
指標PSNR (dB)
資料集與序列FAST-LIVO · f0 hku2
表格設定(擷取紀錄原文):Ablation on sequence f0 (Lang et al., 2025, Table III)
PSNR (dB),FAST-LIVO · f0 hku2
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 Lang 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(Lang et al., 2025, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Gaussian-LIC w/o exposure modelling本方法原文提出 | 29.77 dB | (Lang et al., 2025, Table III) |
| Gaussian-LIC w/o sky modelling本方法原文提出 | 29.76 dB | (Lang et al., 2025, Table III) |
| Gaussian-LIC w/o visual SFM points本方法原文提出 | 29.7 dB | (Lang et al., 2025, Table III) |
| Gaussian-LIC full本方法原文提出 | 29.89 dB | (Lang et al., 2025, Table III) |
Xie et al., 2025 · ICCV Supp. Table 6 本方法 1 筆
指標Mapping FPS
資料集與序列not stated · not stated
表格設定(擷取紀錄原文):Mapping FPS; sequence not stated; Gaussian-LIC value copied from its preprint (marked *), which the GS-LIVM text says ran on an RTX 3090; added by second checker from the ICCV supplement (Xie et al., 2025, ICCV Supp. Table 6)
Mapping FPS,not stated · not stated
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 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 6)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| NeRF-SLAM硬體:not stated | 3.1 fps | (Xie et al., 2025, ICCV Supp. Table 6) |
| MonoGS硬體:not stated | 5.3 fps | (Xie et al., 2025, ICCV Supp. Table 6) |
| 3DGS硬體:not stated | 無數值未報告註記(擷取紀錄):not reported ('-' in table) | (Xie et al., 2025, ICCV Supp. Table 6) |
| Gaussian-LIC* (preprint result)本方法硬體:RTX 3090 GPU (as stated in the GS-LIVM supplement text) | 10 fps | (Xie et al., 2025, ICCV Supp. Table 6) |
| Ours原文提出硬體:Intel Core i9-13900HX (5.50 GHz), 64 GB RAM, NVIDIA RTX 4060 Laptop 8 GB GPU | 12.56 fps | (Xie et al., 2025, ICCV Supp. Table 6) |
來源
Lang et al., 2025
(2025)Gaussian-LIC: Real-Time Photo-Realistic SLAM with Gaussian Splatting and LiDAR-Inertial-Camera Fusion2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 8500-8507
DOI 10.1109/icra55743.2025.11128712arXiv 2404.06926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:2404.06926 (v3, 2025-08-20; latest HTML read) https://arxiv.org/abs/2404.06926
- 後續延伸版:Gaussian-LIC2: LiDAR-Inertial-Camera Gaussian Splatting SLAM (arXiv:2507.04004; IJRR 2026 acceptance claimed only in the repository README, unverified) https://arxiv.org/abs/2507.04004
程式碼:https://github.com/APRIL-ZJU/Gaussian-LIC。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。