Adaptive robust kernels
作者以 Barron 的一般化穩健損失為基礎,把控制核形狀的參數 α 視為未知數,以交替最小化求解:先以一維格點搜尋在 [-10, 2] 內取殘差負對數概似最小的 α,再以迭代重加權最小平方法求解模型參數。為讓 α 可取負值以強力壓低離群值,作者把配分函數的積分截斷在 ±τ(τ = 10c),並預先建立解析度 0.1 的查找表。方法整合進 SuMa 的點對平面投影式 ICP,在 KITTI 里程計序列上不需人工離群值剔除即得到最低的平均平移誤差,並在 CARLA 模擬影像的光束法平差中擴大收斂範圍。
本頁內容
Uses a generalized robust kernel family whose shape is tuned automatically from residuals, tested in ICP and bundle adjustment.
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 scans of the KITTI odometry benchmark (scanner model not named in the paper)、monocular camera images simulated in CARLA (car front-looking, UAV nadir, strong shadows, side-looking with motion blur) |
|---|---|
| 原文測試平台 | car (KITTI dataset)、simulation (CARLA: car-mounted and UAV cameras) |
| 狀態估計 | alternating minimization: (1) alpha chosen by 1-D grid search over [-10, 2] minimizing the negative log-likelihood of the current residuals under a truncated partition function (tau = 10c, lookup table at 0.1 resolution); (2) model parameters by IRLS with the Barron general kernel at that alpha; scale c fixed a priori |
| 資料關聯 | not part of the method; in the ICP experiment it runs inside SuMa frame-to-frame point-to-plane projective ICP without any outlier rejection step; in BA the initial matches come from SIFT with 5-point RANSAC |
| 時間表示 | 不適用 |
| 去畸變 | 不適用 |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | 不適用 |
| 先驗資訊 | none |
| 可輸出幾何 | robust estimate (ICP pose or BA solution) |
| 計算需求 | no runtime or hardware reported; the truncated partition function is precomputed as a lookup table for efficiency |
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 | 3D LiDAR (KITTI odometry scans; model not named in the paper) | 資料集感測器 | KITTI odometry | 原文未報告 | (Chebrolu et al., 2021, Sec. IV-A) |
| 相機 | simulated monocular camera (CARLA simulator) | 資料集感測器 | authors' CARLA-simulated BA datasets | four authors' datasets: car front-looking, UAV nadir, strong shadows, side-looking with motion blur | (Chebrolu et al., 2021, Sec. IV-B, Fig. 6) |
作者報告的優勢與限制
優勢
- best average KITTI relative translational error 2.03% vs 2.35% (Barron), 2.27% (Geman-McClure), 2.90% (SuMa hand-crafted rejection) and 6.34% (Huber) (Table I)
- on sequence 04 Huber (49%) and the hand-crafted scheme (11.9%) fail while the adaptive kernel gives 0.95% (Table I, Sec. IV-A)
- no hand-crafted outlier rejection needed; alpha becomes negative with moving cars on sequence 01 (Sec. IV-A, Fig. 5)
- BA convergence rate 45% vs 24.8% (squared), 33% (Huber), 28.2% (Geman-McClure) (Sec. IV-B)
- BA translation and rotation errors 2 to 5 times lower than with the Huber kernel depending on the dataset, as stated in the text (Sec. IV-B, Fig. 7)
- results insensitive to tau in {10c, 20c, 50c, 100c}: at most about 5% translation and 8% rotation difference (Sec. IV-C)
限制
- not the best in relative rotational error (average 1.18 deg per 100 m vs 0.9 for the hand-crafted scheme and 0.92 for Huber) (Table I, Sec. IV-A)
- scale parameter c is fixed; joint adaptation of alpha and c is unresolved (Sec. V)
- a single alpha per scan pair or per BA iteration; per-object or per-block alpha is future work (Sec. V)
- Agamennoni's elliptical-kernel method outperforms it in translation on the second (UAV) BA dataset (Sec. IV-B)
營建工程相關證據
原文未報告
原文驗證環境:公開基準、模擬
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 25 筆紀錄。
Chebrolu et al., 2021 · Table I 本方法 24 筆
表格設定(擷取紀錄原文):Frame-to-frame point-to-plane projective ICP inside SuMa on KITTI odometry; only the robust kernel or outlier scheme differs between rows (Chebrolu et al., 2021, Table I)
relative rot. error in degrees per 100 m,KITTI odometry · 00
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 Chebrolu et al., 2021 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(Chebrolu et al., 2021, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Our Approach本方法原文提出 | 1.5 deg/100 m | (Chebrolu et al., 2021, Table I) |
| Adaptive Kernel (Barron [6]) | 1.6 deg/100 m | (Chebrolu et al., 2021, Table I) |
| Fixed Kernel (Huber) | 0.93 deg/100 m | (Chebrolu et al., 2021, Table I) |
| Fixed Kernel (Geman-McClure) | 1.8 deg/100 m | (Chebrolu et al., 2021, Table I) |
| Hand-Crafted Outlier Rejection [7] (SuMa original: Huber + rejection of correspondences >2 m or normal angle >30 deg) | 0.9 deg/100 m | (Chebrolu et al., 2021, Table I) |
Chebrolu et al., 2021 · Text Sec. IV-B 本方法 1 筆
指標successful convergence rate
資料集與序列authors' CARLA-simulated bundle adjustment datasets · all 500 instances
表格設定(擷取紀錄原文):BA convergence test: camera poses perturbed with sigma in [0.1 m, 5 m], 20 instances per noise level, 500 instances; converged if final camera-centre RMS error < 1 cm (Chebrolu et al., 2021, Text Sec. IV-B)
successful convergence rate,authors' CARLA-simulated bundle adjustment datasets · all 500 instances
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 Chebrolu et al., 2021 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(Chebrolu et al., 2021, Text Sec. IV-B)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Our approach (adaptive truncated kernel)本方法原文提出 | 45% | (Chebrolu et al., 2021, Sec. IV-B (text)) |
| squared loss | 24.8% | (Chebrolu et al., 2021, Sec. IV-B (text)) |
| Huber | 33% | (Chebrolu et al., 2021, Sec. IV-B (text)) |
| Geman-McClure | 28.2% | (Chebrolu et al., 2021, Sec. IV-B (text)) |
來源
Chebrolu et al., 2021
(2021)Adaptive Robust Kernels for Non-Linear Least Squares ProblemsIEEE Robotics and Automation Letters, 6(2):2240-2247
DOI 10.1109/lra.2021.3061331arXiv 2004.14938
同儕審查已出版已讀全文近十年
相關版本
- 預印本:arXiv:2004.14938 (v1 2020-04-30; v3 2021-02-18) https://arxiv.org/abs/2004.14938