KISS-Matcher
KISS-Matcher 從整體流程角度重新設計全域點雲配準，組合幾何抑制（如地面分割）、改良自 FPFH 的 Faster-PFH 特徵、以 k-core 為基礎的圖論離群剔除（降低 TEASER++ 最大團搜尋的時間複雜度）與 GNC 求解器，並釋出開源 C++ 函式庫。作者在 KITTI、MulRan 迴圈閉合基準以及以 FAST-LIO2 產生的多機器人地圖層級點雲上測試，報告精度與先進方法相當但速度大幅提升，可由掃描擴展到地圖層級。
本頁內容
KISS-Matcher combines Faster-PFH, k-core outlier pruning and a GNC solver into an open-source global registration pipeline scaling from scans to maps.
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 | 64-channel 3D LiDAR (KITTI and MulRan, different ray patterns; models not named in the paper)、map clouds produced by FAST-LIO2-based SLAM on the Kimera-Multi dataset (sensors not stated in the paper) |
|---|---|
| 原文測試平台 | vehicle (KITTI, MulRan)、multi-robot maps from the Kimera-Multi dataset (robot type not stated in the paper) |
| 狀態估計 | maximum k-core pruning of a pairwise-invariant compatibility graph (O(|V|+|E|), CSR storage, beta = 1.5v) followed by a GNC non-minimal solver; the final inlier count is used to reject failed registrations; all parameters scale with voxel size v (r_normal = 3.5v, r_FPFH = 5.0v) |
| 資料關聯 | geometric suppression (Patchwork ground segmentation in the experiments); Faster-PFH with one radius search per point, a linearity filter (tau_lin = 0.99) and minimum neighbour count (tau_num = 3); mutual (reciprocity) matching; top N_tau = 3,000 correspondences by descriptor distance ratio |
| 時間表示 | 不適用 |
| 去畸變 | 不適用 |
| 迴圈閉合 | evaluated on the loop-closing benchmark of Lim et al. (Sec. IV-A) |
| 全域最佳化 | none |
| 地圖表示 | scan-, submap- and map-level point clouds (voxelized) (Sec. IV-A) |
| 先驗資訊 | none (global registration without initial guess) |
| 可輸出幾何 | rigid transformation from scan to map level |
| 計算需求 | entire pipeline about 14 Hz on an Intel Core i9-13900 versus about 6 Hz for other outlier-robust pipelines and about 0.1 Hz (9.57 s per pair) for Predator (Table I caption); Faster-PFH about 4.5x (single-thread) and 2.4x (multi-thread) faster than FPFH (Sec. III-C); more than 20x faster than the TEASER++ pipeline above 200K voxelized points (Sec. IV-E, Fig. 1(b)) |
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 | 64-channel LiDAR sensor (model not named in the paper) | 資料集感測器 | KITTI | laser ray pattern differs from MulRan's | (Lim et al., 2025, Sec. IV-D; Fig. 5 caption) |
| LiDAR | 64-channel LiDAR sensor (model not named in the paper) | 資料集感測器 | MulRan | laser ray pattern differs from KITTI's | (Lim et al., 2025, Sec. IV-D; Fig. 5 caption) |
| 運算硬體 | Intel Core i9-13900 | 執行運算平台 | 未標示 | runtime and parameter studies | (Lim et al., 2025, Table I caption; Sec. III-C; Fig. 7 caption) |
作者報告的優勢與限制
優勢
- on par with state of the art on KITTI while much faster (Sec. V)
- generalizes across datasets and scales without training (Sec. I, IV)
- final inlier count allows rejecting failure cases (Fig. 5 caption)
- 100% success on the KITTI 10 m benchmark, alone and with G-ICP (Table I)
- only compared method that succeeded in map-level registration of Kimera-Multi FAST-LIO2 maps (Sec. IV-D, Fig. 6)
限制
- evaluation centred on driving and multi-robot outdoor datasets; indoor or construction data not reported in sections read (reviewer observation)
- application to mapping/localization left for future work (Sec. V)
- standalone RTE 18.10 cm and RRE 0.94 deg are higher than most compared methods; G-ICP fine alignment lowers RTE to 1.10 cm (Table I)
- runtime still grows linearly with the number of correspondences, hence the N_tau cap (Sec. III-D)
- (reviewer observation) map-level evaluation is qualitative only (Fig. 6)
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 4 個比較組，合計 10 筆紀錄。
Lim et al., 2025 · Table I 本方法 6 筆
資料集與序列KITTI · 10 m benchmark
表格設定（擷取紀錄原文）：KITTI 10 m benchmark [23]: scan-to-scan global registration; success if translation < 2 m and rotation < 5 deg; RTE and RRE averaged over successful registrations only (Sec. IV-A); W = submap window size; learning-based results as listed by the authors (Lim et al., 2025, Table I)
RTE [cm]，KITTI · 10 m benchmark
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 Lim 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(Lim et al., 2025, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| 3DFeat-Net | 25.9 cm | (Lim et al., 2025, Table I) |
| FCGF | 6.47 cm | (Lim et al., 2025, Table I) |
| DIP | 8.69 cm | (Lim et al., 2025, Table I) |
| Predator | 5.6 cm | (Lim et al., 2025, Table I) |
| SpinNet | 9.88 cm | (Lim et al., 2025, Table I) |
| D3Feat | 11 cm | (Lim et al., 2025, Table I) |
| GeDi | 7.55 cm | (Lim et al., 2025, Table I) |
| G-ICP | 8.56 cm | (Lim et al., 2025, Table I) |
| STD, W = 1 | 26.09 cm | (Lim et al., 2025, Table I) |
| STD, W = 3 | 20.94 cm | (Lim et al., 2025, Table I) |
| STD, W = 5 | 23.97 cm | (Lim et al., 2025, Table I) |
| MapClosures, W = 1 | 38.5 cm | (Lim et al., 2025, Table I) |
| MapClosures, W = 3 | 31.5 cm | (Lim et al., 2025, Table I) |
| MapClosures, W = 5 | 32.27 cm | (Lim et al., 2025, Table I) |
| FPFH + FGR | 6.94 cm | (Lim et al., 2025, Table I) |
| FPFH + TEASER++ | 9.36 cm | (Lim et al., 2025, Table I) |
| FPFH + Quatro | 13.15 cm | (Lim et al., 2025, Table I) |
| Proposed本方法原文提出 | 18.1 cm | (Lim et al., 2025, Table I) |
| FPFH + FGR + G-ICP | 1.22 cm | (Lim et al., 2025, Table I) |
| FPFH + TEASER + G-ICP | 1.1 cm | (Lim et al., 2025, Table I) |
| FPFH + Quatro + G-ICP | 1.1 cm | (Lim et al., 2025, Table I) |
| Proposed + G-ICP本方法原文提出 | 1.1 cm | (Lim et al., 2025, Table I) |
Lim et al., 2025 · Text Sec. III-C 本方法 2 筆
資料集與序列KITTI and MulRan · 2-12 m loop closing test
表格設定（擷取紀錄原文）：speed-up of Faster-PFH over FPFH while maintaining performance (Lim et al., 2025, Text Sec. III-C)
speed improved approximately 4.5 times (single-threaded)，KITTI and MulRan · 2-12 m loop closing test
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 Lim 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 |
|---|---|---|
| Faster-PFH vs FPFH本方法原文提出硬體：not stated for these ratios (Fig. 8 caption names no CPU; Sec. III-C and Fig. 7 name an Intel Core i9-13900) | 4.5 x | (Lim et al., 2025, Sec. III-C; Fig. 8(a) caption) |
Lim et al., 2025 · Text Sec. IV-E 本方法 1 筆
指標more than a 20x speed improvement
資料集與序列not named (large-scale registration at kilometre level, Fig. 1(b)) · voxelized clouds above 200K points
表格設定（擷取紀錄原文）：speed-up of the entire pipeline over the full TEASER++ pipeline in large-scale registration at the kilometre level (Sec. IV-E, Fig. 1(b)); the dataset behind Fig. 1(b) is not named (Lim et al., 2025, Text Sec. IV-E)
more than a 20x speed improvement，not named (large-scale registration at kilometre level, Fig. 1(b)) · voxelized clouds above 200K points
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 Lim 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 |
|---|---|---|
| KISS-Matcher vs TEASER++ pipeline本方法原文提出 | 20 x僅報告範圍註記（擷取紀錄）：lower bound (more than 20x) | (Lim et al., 2025, Sec. IV-E; Fig. 1(b)) |
Lim et al., 2025 · Text Table I caption 本方法 1 筆
指標operates around 14 Hz for the entire pipeline
資料集與序列KITTI · 10 m benchmark
表格設定（擷取紀錄原文）：entire pipeline rate (feature extraction and matching to pose estimation) on the KITTI 10 m benchmark, stated in the Table I caption (Lim et al., 2025, Text Table I caption)
operates around 14 Hz for the entire pipeline，KITTI · 10 m benchmark
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 Lim 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 |
|---|---|---|
| KISS-Matcher本方法原文提出硬體：Intel Core i9-13900 CPU | 14 Hz | (Lim et al., 2025, Table I caption) |
來源
Lim et al., 2025
（2025）KISS-Matcher: Fast and Robust Point Cloud Registration Revisited2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 11104-11111
DOI 10.1109/icra55743.2025.11127458arXiv 2409.15615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:2409.15615 (v1 2024-09-23; v3 2025-07-16) https://arxiv.org/abs/2409.15615
- 程式碼釋出：KISS-Matcher https://github.com/MIT-SPARK/KISS-Matcher
程式碼：https://github.com/MIT-SPARK/KISS-Matcher（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。