TEASER / TEASER++
TEASER 以截斷最小平方(Truncated Least Squares, TLS)成本處理大量錯誤對應,並以旋轉平移不變量的圖論框架將尺度、旋轉與平移分解後依序求解。尺度與平移以自適應投票求解,旋轉以半正定鬆弛(TEASER)或逐步非凸化(TEASER++)求解,並用最大團剔除大量離群對應。TEASER++ 另以 Douglas-Rachford 分裂計算最佳性證書,可用來辨識不可靠的配準結果。在 3DMatch 以 3DSmoothNet 對應點測試時,TEASER++ 平均 0.059 秒,八個場景中除 MIT Lab 外成功率都不低於 RANSAC-10K;經證書篩選(CERT)後成功率更高,但證書計算平均需 238 秒。場景對稱或正確對應少於 3 組時仍會失敗。
本頁內容
Certifiable robust registration via TLS cost, decoupled scale/rotation/translation, max-clique outlier pruning, and GNC plus Douglas-Rachford certification in TEASER++.
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 | ["RGB-D (3DMatch scans in experiments)", "RGB-D (large-scale hierarchical multi-view RGB-D object dataset [36], object pose tests)", "sensor-agnostic 3D correspondences"] |
|---|---|
| 原文測試平台 | 未記錄 |
| 狀態估計 | truncated least squares; adaptive voting for scale and translation; max-clique pruning on invariant measurements; SDP relaxation (TEASER) or GNC with Douglas-Rachford certification (TEASER++) for rotation (abstract, Sec. X) |
| 資料關聯 | putative correspondences from FPFH for object pose (Sec. XI-D) or 3DSmoothNet descriptors with nearest-neighbour matching on the 5,000 provided keypoints per 3DMatch scan (Sec. XI-E); all-to-all hypotheses (|A| x |B| about 10^4 for 100-point clouds) in the correspondence-free test (Sec. XI-C) |
| 時間表示 | 不適用 |
| 去畸變 | 不適用 |
| 迴圈閉合 | authors suggest certified registration for loop-closure validation in SLAM (Sec. XI) |
| 全域最佳化 | certifiable global optimality of the rotation subproblem (abstract) |
| 地圖表示 | 3D point correspondences |
| 先驗資訊 | none (no initial guess) |
| 可輸出幾何 | scale, rotation, translation with optimality certificate |
| 計算需求 | TEASER in MATLAB with cvx for the SDP is impractical for real time; TEASER++ in C++ (Eigen, OpenMP, parallel max clique) runs under 10 ms with known scale and under 30 ms with unknown scale for N = 100 on a laptop with an i7-8850H CPU and 32 GB RAM (Sec. XI, XI-B); about 2 s for 10,000 correspondences at 95% outliers (App. S); on 3DMatch, 0.059 s average vs 0.008 s for RANSAC-1K and 0.074 s for RANSAC-10K on a Xeon Platinum 8259CL at 2.50 GHz with 12 threads (Table II); certification averages 238.136 s on 3DMatch, 24 DRS iterations of 50 ms each in the synthetic test, and about 1200 s for the full SDP with K = 100 in MOSEK (Sec. VIII-C, XI-A, Table II) |
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 |
|---|---|---|---|---|---|
| RGB-D 相機 | 原文未報告 | 資料集感測器 | 3DMatch | RGB-D scans of 62 indoor scenes (54 train, 8 test); 5,000 keypoints per scan provided | (Yang et al., 2021, Sec. XI-E) |
| RGB-D 相機 | 原文未報告 | 資料集感測器 | large-scale hierarchical multi-view RGB-D object dataset [36] | large-scale RGB-D point cloud scenes with object labels | (Yang et al., 2021, Sec. XI-D, App. T) |
| 運算硬體 | i7-8850H CPU, 32GB RAM (laptop) | 執行運算平台 | 未標示 | all tests except 3DMatch | (Yang et al., 2021, Sec. XI (implementation details)) |
| 運算硬體 | Xeon Platinum 8259CL CPU at 2.50GHz (server) | 執行運算平台 | 未標示 | server CPU; 12 threads allocated per algorithm; used for the 3DMatch scan-matching tests | (Yang et al., 2021, Sec. XI-E) |
作者報告的優勢與限制
優勢
- robust to more than 99% outliers with known scale on benchmarks, with lower errors than GORE and about one order of magnitude faster (Sec. XI-B)
- TEASER++ runs in milliseconds (abstract, Sec. XI-B)
- MCIS pruning reduces outlier rates to below 10% (Sec. XI-A)
- on 3DMatch TEASER++ success is 83.1% to 98.6% per scene and not below RANSAC-10K except MIT Lab, and certified results reach 88.6% to 99.4% (Table II)
- object pose on eight RGB-D scenes with mean FPFH inlier ratio 6.53% gives mean errors 0.066 rad and 0.069 m (Table I)
- correspondence-free registration succeeds down to 10% overlap and is more accurate than Go-ICP (Sec. XI-C)
限制
- fails on scenes with symmetric keypoint distributions, e.g., a 180-degree wrong solution in 3DMatch Hotel 3 (Fig. 11)
- cannot succeed when fewer than 3 inliers exist, which is not uncommon with current descriptors (Sec. XI-E)
- adversarial outliers make outlier rejection ill-posed (Sec. IX, estimation contract)
- with unknown scale TEASER and TEASER++ fail at 90% outliers (Sec. XI-B)
- the GNC rotation solver alone fails above about 80% outliers, so TEASER++ depends on MCIS pruning (Sec. XI-A, X)
- the correspondence-free mode needs quadratically many hypotheses, is dominated by max-clique and scale estimation time and fails below 10% overlap (Sec. XI-C)
- TEASER with the full SDP is impractical for real time (Sec. XI-B)
- RANSAC slightly better on the Lab scene of 3DMatch (Sec. XI)
- running the certification (TEASER++ CERT) on 3DMatch required more than 200 s on average, although most instances were certified within 100 s (Sec. XI)
營建工程相關證據
原文未報告(對稱性造成失敗的案例與建物重複結構相關,屬推論)
原文驗證環境:模擬、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 11 個比較組,合計 84 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 7 組列在最後,並連到性能比較頁。
Yang et al., 2021 · Text App. T 本方法 24 筆
表格設定(擷取紀錄原文):Per-scene values listed under the Fig. 18 panels in App. T for the Table I experiment; units are not printed per scene (Table I uses rad and m); per-scene object label (cereal box or cap) not transcribed (Yang et al., 2021, Text App. T)
Rotation error,large-scale hierarchical multi-view RGB-D object dataset [36] · scene-1
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 Yang et al., 2021 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 |
|---|---|---|
| TEASER本方法原文提出 | 0.066 rad | (Yang et al., 2021, App. T, Fig. 18) |
Yang et al., 2021 · Table II 本方法 18 筆
表格設定(擷取紀錄原文,這些數值分屬表中不同部分):(Yang et al., 2021, Table II)
- 3DMatch test scenes, 3DSmoothNet correspondences, success = rotation error < 10 deg and translation error < 30 cm; beta = 5 cm; CERT = subset certified with sub-optimality gap 3%; Xeon Platinum 8259CL, 12 threads
- Same setting as other Table II rows
percentage of correct registration results,3DMatch · Kitchen
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 Yang 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(Yang et al., 2021, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| RANSAC-1K | 91.3% | (Yang et al., 2021, Table II) |
| RANSAC-10K | 97.2% | (Yang et al., 2021, Table II) |
| TEASER++本方法原文提出 | 98.6% | (Yang et al., 2021, Table II) |
| TEASER++ (CERT)本方法原文提出 | 99.4% | (Yang et al., 2021, Table II) |
Chang et al., 2022 · Table II 本方法 16 筆
表格設定(擷取紀錄原文):Loop-closure relative pose estimation with different ICP initializations on ground-truth and false loop-closure sets; SAC cumulative error threshold 32 m (500 iterations), ICP threshold 0.9 m (200 iterations); errors computed on correct loop closures that passed SAC and ICP (Chang et al., 2022, Table II)
recall of correct loop closures passing SAC and ICP,CoSTAR multi-robot dataset: Tunnel · Tunnel
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 Chang et al., 2022 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(Chang et al., 2022, Table II)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| GT initialization (oracle) | 90.8% | (Chang et al., 2022, Table II) |
| OdomRot [8] initialization (LAMP 1.0) | 93.9% | (Chang et al., 2022, Table II) |
| TEASER++ initialization + GICP本方法原文提出 | 76.6% | (Chang et al., 2022, Table II) |
| SAC-IA initialization + GICP原文提出 | 81.9% | (Chang et al., 2022, Table II) |
Yang et al., 2021 · Table I 本方法 8 筆
資料集與序列large-scale hierarchical multi-view RGB-D object dataset [36] · eight scenes
表格設定(擷取紀錄原文):Object pose estimation on eight scenes of the UW RGB-D dataset [36]: object (cereal box or cap) cut from the scene, scene randomly transformed, FPFH correspondences, both downsampled so the object has 2,000 points; mean and SD over the eight scenes (Yang et al., 2021, Table I)
Rotation error [rad],large-scale hierarchical multi-view RGB-D object dataset [36] · eight scenes
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 Yang et al., 2021 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 |
|---|---|---|
| TEASER本方法原文提出 | 0.066 rad | (Yang et al., 2021, Table I) |
其他比較組
來源
相關版本
- 預印本:arXiv:2001.07715 (v1 2020-01-21; v2 2020-10-17) https://arxiv.org/abs/2001.07715
- 程式碼釋出:TEASER-plusplus https://github.com/MIT-SPARK/TEASER-plusplus
程式碼:https://github.com/MIT-SPARK/TEASER-plusplus(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。