LIT-GS
LIT-GS 以熱影像取代易受光照影響的 RGB 光度監督,建立光達、慣性與熱影像的高斯潑濺(Gaussian Splatting)地圖。它以上游 FAST-LIVO2 的具不確定度視覺地圖點作為跨模態錨點建立熱影像與光達的對應,並把加權的光達點到平面殘差放入光束法平差,再於高斯最佳化中加入光達平面正則化,以抑制表面增厚與結構漂移。此方法為離線流程,作者僅報告 Car 場景的訓練時間為 70.6 分鐘。
本頁內容
An offline LiDAR-inertial-thermal Gaussian splatting pipeline that anchors thermal-LiDAR associations on FAST-LIVO2 map points and injects LiDAR plane constraints into both bundle adjustment and Gaussian training.
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 Avia, 10 Hz per Fig. 2)、IMU (built into the Livox Avia, 200 Hz per Fig. 2)、visible-light camera (MV-CA013-21UC)、long-wave thermal imager (MV-CI003-GL-N15, 10 Hz per Fig. 2) |
|---|---|
| 原文測試平台 | handheld (inferred: the paper uses a modified version of a sensor suite whose repository is named LIV_handhold, and Fig. 1 shows battery, on-board PC and display; the carrying mode is not stated in the text) |
| 狀態估計 | upstream FAST-LIVO2 LIV estimate, then offline LiDAR-plane-constrained bundle adjustment (extension of COLMAP-PCD) and differentiable Gaussian optimization |
| 資料關聯 | learned thermal feature matching; uncertainty-tagged LIV visual map points as cross-modal anchors; weighted LiDAR point-to-plane residuals in BA and as a splatting regularizer |
| 時間表示 | 原文未報告 (PPS triggers from a microcontroller synchronize LiDAR, IMU and thermal camera; Fig. 1 a3 labels an STM32 board and GPRMC over USB serial) |
| 去畸變 | 原文未報告 |
| 迴圈閉合 | none reported |
| 全域最佳化 | offline bundle adjustment over poses and triangulated points |
| 地圖表示 | 3D Gaussians (thermal), initialized from LiDAR voxel map and refined structure |
| 先驗資訊 | upstream FAST-LIVO2 map points and poses |
| 可輸出幾何 | thermal Gaussian map with rendered images; geometric consistency evaluated with Earth Mover's Distance to reference point clouds (Sec. IV-B2) |
| 計算需求 | offline: training on one scene 70.6 min; rendering about 25.3 ms per 640x512 frame on i7-14700KF with RTX 4080D (Sec. IV-C) |
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 | 方法輸入 | 未標示 | built-in IMU; 10 Hz (Fig. 2) | (Shi et al., 2026, Sec. IV-A; Fig. 1; Fig. 2) |
| 慣性量測單元(IMU) | built-in IMU of the Livox Avia | 方法輸入 | 未標示 | 200 Hz (Fig. 2) | (Shi et al., 2026, Sec. IV-A; Fig. 2) |
| 相機 | MV-CA013-21UC visible-light camera | 資料集感測器 | authors' private thermal dataset (self-collected rig) | 原文未報告 | (Shi et al., 2026, Sec. IV-A) |
| 熱像儀 | MV-CI003-GL-N15 long-wave thermal imager | 方法輸入 | 未標示 | 10 Hz (Fig. 2); rendering evaluated at 640 x 512 per frame (Sec. IV-C); native imager resolution not stated | (Shi et al., 2026, Sec. IV-A; Sec. IV-C; Fig. 2) |
| 運算硬體 | Intel Core i7-14700KF CPU | 執行運算平台 | 未標示 | 原文未報告 | (Shi et al., 2026, Sec. IV-A) |
| 運算硬體 | NVIDIA GeForce RTX 4080D GPU | 執行運算平台 | 未標示 | 原文未報告 | (Shi et al., 2026, Sec. IV-A) |
| 其他 | STM32 microcontroller (PPS synchronization board) | 方法輸入 | 未標示 | synchronized timers, GPRMC over USB serial (Fig. 1 a3) | (Shi et al., 2026, Sec. III; Fig. 1 (a3)) |
作者報告的優勢與限制
優勢
- Lower EMD than LIV-GaussMap on private and public scenes, notably in plane-dominated scenes (Sec. IV-B2, Table I)
- Illumination-robust supervision from thermal imagery in low-light scenes (Sec. I, V)
限制
- RGB-LiDAR Gaussian mapping outperforms thermal-LiDAR under good lighting (Sec. V)
- Offline training time of 70.6 min for one scene (Sec. IV-C)
- Residual thermal-LiDAR extrinsic errors require refinement (Sec. III-B)
- Thermal3D-GS has a lower average EMD than LIT-GS on the five private scenes (0.179 vs 0.191; also lower on Sundial, Tree-stump and Bushes), although the text says its EMD values are generally higher (Table I; Sec. IV-B2)
- Source and accuracy of the reference point clouds used for EMD are not described (Sec. IV-B2)
營建工程相關證據
原文未報告(私人資料為 Car、Logo、Sundial、Tree-stump、Bushes 五個場景;公開資料為 M2DGR 的 Street-03、Gate-01、Lift-01、Door-02、Hall-02;未見工地)。熱影像與幾何正則化對夜間或低照度隧道檢測可能有參考價值(推論)。
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 2 個比較組,合計 50 筆紀錄。
Shi et al., 2026 · Table I 本方法 48 筆
表格設定(擷取紀錄原文):Thermal Gaussian-splatting reconstruction quality; private scenes captured with the authors' Livox Avia + thermal rig at different times of day (Fig. 3 caption lists 12:00 p.m. to 6:00 a.m.), public scenes are five M2DGR sequences; EMD reference point cloud source not described; Avg. columns printed by the authors (Shi et al., 2026, Table I)
PSNR (rendering fidelity, higher is better),authors' private thermal dataset · Car
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 Shi et al., 2026 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(Shi et al., 2026, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Ours本方法原文提出 | 26.061 | (Shi et al., 2026, Table I) |
| Thermal3D-GS | 24.216 | (Shi et al., 2026, Table I) |
| LIV-GaussMap | 23.569 | (Shi et al., 2026, Table I) |
Shi et al., 2026 · Text Sec.IV-C 本方法 2 筆
資料集與序列authors' private thermal dataset · Car
表格設定(擷取紀錄原文):Offline training and rendering cost (Shi et al., 2026, Text Sec.IV-C)
training time on the Car scene,authors' private thermal dataset · Car
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 Shi et al., 2026 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 |
|---|---|---|
| LIT-GS本方法原文提出硬體:Intel Core i7-14700KF CPU + NVIDIA GeForce RTX 4080D GPU | 70.6 min | (Shi et al., 2026, Sec. IV-C) |
來源
Shi et al., 2026
(2026)LIT-GS: LiDAR-Inertial-Thermal Gaussian Splatting for Illumination-Robust MappingarXiv (author comment: accepted to IEEE/RSJ IROS 2026)
已接受(作者聲明)已讀全文近十年查證後修正
相關版本
- 預印本:LIT-GS arXiv v1 (2026-06-18) https://arxiv.org/abs/2606.20424