KISS-ICP
KISS-ICP 回歸最基本的點到點(point-to-point)ICP,僅保留等速運動預測與逐點去畸變、體素雙重降採樣、依運動模型偏差自適應的對應距離門檻,以及穩健核函數等少數元件。地圖為雜湊表中的降採樣體素點雲,並保留原始點座標以避免離散化誤差。作者主張在同一組參數下可用於車載、無人機、Segway 與手持 LiDAR,且不需 IMU,也不含迴圈或位姿圖。
本頁內容
A minimal LiDAR-only odometry built on robust point-to-point ICP with constant-velocity deskewing, double voxel downsampling, and an adaptive correspondence threshold, working across platforms with one parameter set.
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 only |
|---|---|
| 原文測試平台 | vehicle、UAV、wheeled UGV、handheld |
| 狀態估計 | robust point-to-point ICP (Gauss-Newton with robust kernel) frame-to-local-map, constant-velocity motion prediction |
| 資料關聯 | point-to-point nearest neighbour with adaptive correspondence threshold derived from observed deviation from the motion model |
| 時間表示 | discrete poses with per-point constant-velocity deskew |
| 去畸變 | constant-velocity model applied with per-point relative timestamps (Sec. III-A); IMU or wheel odometry can replace the velocity source |
| 迴圈閉合 | none |
| 全域最佳化 | none (pose-graph optimization stated as orthogonal and not used) |
| 地圖表示 | voxelized, downsampled local point cloud stored in a hash table with a maximum number of points per voxel; voxels beyond maximum range removed |
| 先驗資訊 | none |
| 可輸出幾何 | odometry and local voxel point map; original point coordinates retained within voxels (no centroid snapping) |
| 計算需求 | CPU; faster than sensor frame rate on all presented datasets (abstract; Sec. V); KITTI-raw average 38 Hz with deskewing and 51 Hz without (Table V); computing hardware not stated and detailed runtime deferred to the project page (Sec. IV-A) |
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 | 原文未報告 | 資料集感測器 | handheld Livox data cited as [17] (Loam_livox paper; Fig. 1 only) | handheld Livox LiDAR, model not given; non-repetitive shooting pattern different from rotating mechanical LiDARs | (Vizzo et al., 2023, Fig. 1 caption) |
作者報告的優勢與限制
優勢
- Only seven parameters and the same configuration across datasets (Table I; Sec. IV)
- Constant-velocity deskewing performs on par with IMU-based velocity for deskewing on KITTI-raw in their test (Sec. IV-D; Table V)
- Lowest relative and absolute errors on all MulRan sequences versus MULLS, SuMa and F-LOAM, e.g. KAIST ATE 17.40 m versus 37.24 m for MULLS (Table III)
- Adaptive threshold matches or beats every fixed threshold on KITTI (average 0.50% versus 0.51 to 0.53%) (Table VI)
限制
- Assumes motion within a sweep is small (Sec. III-A)
- Performance gap on Newer College long sequence attributed to CT-ICP loop closure (Sec. IV-C)
- NCLT evaluation unreliable per the authors (ground-truth misalignment, missing frames); CT-ICP NCLT result could not be reproduced (Sec. IV-C)
- No loop closure or pose-graph optimization; authors describe pose-graph optimization as orthogonal (Sec. II)
- Reported to fail on all ENWIDE degenerate sequences (COIN-LIO, Sec. IV-C)
- Reported to struggle with quadruped and backpack motion under constant-velocity model (RKO-LIO, Sec. II)
營建工程相關證據
原文未報告
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 44 個比較組,合計 242 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 40 組列在最後,並連到性能比較頁。
Ferrari et al., 2024 · Table II 本方法 19 筆
表格設定(擷取紀錄原文):KITTI benchmark RPE (%); segments 100-800 m for KITTI, MulRan and NC1, 10-80 m for NC0 and Hilti; averages exclude failures; per-sequence KITTI 00-10 and MulRan rows omitted (averages kept); values identical in arXiv v1 and version of record (Ferrari et al., 2024, Table II)
RPE [%] (segments 10-80 m),Newer College NC0 (OS0-128) · cat. easy
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 Ferrari et al., 2024 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(Ferrari et al., 2024, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Ours (MAD-ICP)原文提出 | 1.16% | (Ferrari et al., 2024, Table II) |
| KISS-ICP本方法 | 2.06% | (Ferrari et al., 2024, Table II) |
| F-LOAM | 1.36% | (Ferrari et al., 2024, Table II) |
| MULLS | 2% | (Ferrari et al., 2024, Table II) |
| CT-ICP | 1.12% | (Ferrari et al., 2024, Table II) |
Vizzo et al., 2023 · Table III 本方法 16 筆
表格設定(擷取紀錄原文):MulRan; values are averages over the three runs per sequence; CT-ICP not evaluated because it lacks MulRan support (Vizzo et al., 2023, Table III)
Avg. tra. (KITTI relative translational error),MulRan · KAIST
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 Vizzo et al., 2023 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(Vizzo et al., 2023, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MULLS [21] | 2.94% | (Vizzo et al., 2023, Table III) |
| SuMa [1] | 5.59% | (Vizzo et al., 2023, Table III) |
| F-LOAM [33] | 3.43% | (Vizzo et al., 2023, Table III) |
| Ours (KISS-ICP)本方法原文提出 | 2.28% | (Vizzo et al., 2023, Table III) |
Guadagnino et al., 2025b · Table II 本方法 14 筆
表格設定(擷取紀錄原文):RPE is the KITTI average translation error over 1, 2, 5, 10, 20, 50 and 100 m segments (%); ATE is RMS absolute translation error after alignment (m); warehouse reference is Cartographer, campus and park reference is the total station (Guadagnino et al., 2025b, Table II)
RPE [%] (KITTI metric, 1-100 m segments),authors' warehouse and campus sequences · Campus 0
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 Guadagnino et al., 2025b 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(Guadagnino et al., 2025b, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Wheel Odometry | 4.93% | (Guadagnino et al., 2025b, Table II) |
| KISS-ICP [31]本方法 | 4.9% | (Guadagnino et al., 2025b, Table II) |
| WO + 3D KISS-ICP | 4.64% | (Guadagnino et al., 2025b, Table II) |
| WO + 2D KISS-ICP | 4.43% | (Guadagnino et al., 2025b, Table II) |
| EKF (robot_localization fusing WO + 2D KISS-ICP) | 6.28% | (Guadagnino et al., 2025b, Table II) |
| Fuse (fixed-lag smoother fusing WO + 2D KISS-ICP) | 4.16% | (Guadagnino et al., 2025b, Table II) |
| Kinematic-ICP原文提出 | 2.97% | (Guadagnino et al., 2025b, Table II) |
Blanco-Claraco, 2025 · Table 9 本方法 14 筆
指標absolute translational error (RMSE, evo_ape -a)
表格設定(擷取紀錄原文):Handheld Newer College sequences; ATE RMSE from evo_ape -a (Umeyama alignment); x(value) marks divergence; no method uses the IMU; same default configuration for all datasets (Blanco-Claraco, 2025, Table 9)
absolute translational error (RMSE, evo_ape -a),Newer College (2020, sequences 01 and 02) · 01
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。
- 發散
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 Blanco-Claraco, 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(Blanco-Claraco, 2025, Table 9)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| KISS-ICP本方法 | 0.61 m | (Blanco-Claraco, 2025, Table 9) |
| MOLA-LO (ours)原文提出 | 0.68 m | (Blanco-Claraco, 2025, Table 9) |
| MOLA-LO (always updates local map) | 5.11 m發散註記(擷取紀錄):diverged (value printed in parentheses) | (Blanco-Claraco, 2025, Table 9) |
| MOLA-LO + LC (ours)原文提出 | 0.31 m | (Blanco-Claraco, 2025, Table 9) |
其他比較組
列出其餘 40 個比較組
- Zheng & Zhu, 2024 · Table I
- Zheng & Zhu, 2024 · Table III
- Pfreundschuh et al., 2024 · Table II
- Malladi et al., 2026 · Table I
- Pfreundschuh et al., 2024 · Table I
- Blanco-Claraco, 2025 · Table 4
- Pan et al., 2024 · Table VII
- Pan et al., 2024 · Table VIII
- Malladi et al., 2026 · Table III
- Malladi et al., 2026 · Table IV
- Schillberg et al., 2025 · Table 2
- Schillberg et al., 2025 · Table 3
- Burnett et al., 2025 · Table II
- Zhu et al., 2025 · Table III
- Blanco-Claraco, 2025 · Table 8
- Malladi et al., 2026 · Table II
- Lee et al., 2025a · Table II
- Lee et al., 2025a · Table V
- Guadagnino et al., 2025a · Table II
- Blanco-Claraco, 2025 · Table 10
- Pan et al., 2025 · Table III
- Vizzo et al., 2023 · Table IV
- Vizzo et al., 2023 · Table V
- Vizzo et al., 2023 · Table VI
- Lee et al., 2024b · Table 4
- Blanco-Claraco, 2025 · Table 3
- Burnett et al., 2025 · Table I
- Lee et al., 2025a · Table I
- Lee et al., 2025a · Table VI
- Vizzo et al., 2023 · Table II
- Blanco-Claraco, 2025 · Table 5
- Zheng & Zhu, 2024 · Table V
- Zhu et al., 2025 · Table I
- Zhu et al., 2025 · Table IV
- Lee et al., 2025a · Table III
- Lee et al., 2025a · Table IV
- Ferrari et al., 2024 · Table III
- Blanco-Claraco, 2025 · Table 7
- Blanco-Claraco, 2025 · Text Sec. 4.6
- Zhu et al., 2025 · Table II
來源
Vizzo et al., 2023
(2023)KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right WayIEEE Robotics and Automation Letters, 8(2):1029-1036
DOI 10.1109/lra.2023.3236571arXiv 2209.15397程式碼
同儕審查已出版已讀全文近十年
相關版本
- 預印本:KISS-ICP (arXiv v2) https://arxiv.org/abs/2209.15397
- 程式碼釋出:PRBonn/kiss-icp https://github.com/PRBonn/kiss-icp
程式碼:https://github.com/PRBonn/kiss-icp(授權:MIT (LICENSE file checked))。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。