GeoTransformer
GeoTransformer 屬學習式、免關鍵點的配準:先在降採樣的超點(superpoint)間比對,再傳播到稠密點。其幾何 Transformer 編碼點對距離與三點角度,使特徵對剛體變換不變,並在低重疊情形下保持穩健;摘要指出匹配精度高到不需 RANSAC 即可估計轉換。
本頁內容
Keypoint-free learned registration matching superpoints with a transformation-invariant geometric transformer, removing the need for RANSAC.
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 | none of its own; public benchmarks: 3DMatch and 3DLoMatch indoor RGB-D scene fragments, KITTI odometry LiDAR scans (pairs at least 10 m apart, ground truth refined by ICP), ModelNet40 synthetic CAD points, Augmented ICL-NUIM synthetic RGB-D with a noise model, 4DMatch and 4DLoMatch non-rigid animations |
|---|---|
| 原文測試平台 | 不適用 (offline benchmark evaluation) |
| 狀態估計 | local-to-global registration (LGR): weighted SVD on the point correspondences of each superpoint match gives candidate transforms, the candidate with most inliers within an acceptance radius is kept and re-estimated on inliers for Nr = 5 iterations; RANSAC-50k and plain weighted SVD are also evaluated |
| 資料關聯 | KPConv-FPN backbone; geometric self-attention (pair-wise distance and triplet-wise angle embeddings) interleaved three times with feature-based cross-attention; Gaussian correlation with dual normalisation selects the top Nc superpoint matches (256 at test time); an optimal-transport layer (Sinkhorn) with mutual top-k extracts dense point correspondences inside matched patches |
| 時間表示 | 不適用 |
| 去畸變 | 不適用 |
| 迴圈閉合 | none |
| 全域最佳化 | none for pairwise registration; in the Augmented ICL-NUIM multiway test, pairwise GeoTransformer results are followed by global pose-graph optimisation |
| 地圖表示 | superpoints and dense points |
| 先驗資訊 | trained model per benchmark (40 to 200 epochs); the Augmented ICL-NUIM test reuses the 3DMatch model without fine-tuning |
| 可輸出幾何 | rigid transformation |
| 計算需求 | PyTorch on an RTX 3090 GPU; 3DMatch/3DLoMatch average total time 0.088 s per pair with LGR vs 1.633 s with RANSAC-50k (pose time 0.013 s vs 1.558 s); the lite model runs at 0.073 s, about 13 fps |
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 |
|---|---|---|---|---|---|
| 運算硬體 | RTX 3090 GPU | 執行運算平台 | 未標示 | used for training and testing in PyTorch | (Qin et al., 2023, Sec. 4.1) |
作者報告的優勢與限制
優勢
- improves inlier ratio by 18 to 31 points and registration recall by over 7 points on 3DLoMatch (abstract, Sec. 4.2)
- RANSAC-free LGR: registration recall 91.8% (3DMatch) and 74.5% (3DLoMatch) with 0.088 s average total time vs 1.633 s with RANSAC-50k (Table 2)
- KITTI: RTE 6.8 cm, RRE 0.24 deg, RR 99.8% with LGR (Table 3)
- Augmented ICL-NUIM with the 3DMatch-trained model: mean ATE 14.12 cm, best on three of four trajectories (Table 5)
限制
- learned priors may not transfer across sensors or scenes (inference; cross-dataset degradation reported for a different learned method in Lim et al., 2025 Fig. 5)
- uniformly downsampled superpoints can cause a large memory footprint and computational cost for inputs covering a large area; an extra downsampling stage was needed on KITTI and Augmented ICL-NUIM (Sec. 5)
- uniform superpoint sampling is inflexible and can split a single object into several patches (Sec. 5)
- the backbone is not invariant to rotation, so performance drops under large rotations (Sec. 4.4)
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 6 個比較組,合計 70 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 2 組列在最後,並連到性能比較頁。
Qin et al., 2023 · Table 2 本方法 30 筆
表格設定(擷取紀錄原文):3DMatch (overlap above 30%) and 3DLoMatch (10% to 30%) test pairs; registration recall = share of pairs with transformation RMSE below 0.2 m; model time = feature extraction, pose time = transformation estimation, averaged over all pairs (Qin et al., 2023, Table 2)
Registration Recall RR (%),3DMatch · test split
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 Qin 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(Qin et al., 2023, Table 2)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| FCGF + RANSAC-50k (5000 samples) | 85.1% | (Qin et al., 2023, Table 2) |
| D3Feat + RANSAC-50k (5000 samples) | 81.6% | (Qin et al., 2023, Table 2) |
| SpinNet + RANSAC-50k (5000 samples) | 88.6% | (Qin et al., 2023, Table 2) |
| Predator + RANSAC-50k (5000 samples) | 89% | (Qin et al., 2023, Table 2) |
| CoFiNet + RANSAC-50k (5000 samples) | 89.3% | (Qin et al., 2023, Table 2) |
| GeoTransformer (ours) + RANSAC-50k (5000 samples)本方法原文提出 | 92.3% | (Qin et al., 2023, Table 2) |
| GeoTransformer lite (ours, shared geometric self-attention) + RANSAC-50k (5000 samples)本方法原文提出 | 92.2% | (Qin et al., 2023, Table 2) |
| FCGF + weighted SVD (250 samples) | 42.1% | (Qin et al., 2023, Table 2) |
| D3Feat + weighted SVD (250 samples) | 37.4% | (Qin et al., 2023, Table 2) |
| SpinNet + weighted SVD (250 samples) | 34% | (Qin et al., 2023, Table 2) |
| Predator + weighted SVD (250 samples) | 50% | (Qin et al., 2023, Table 2) |
| CoFiNet + weighted SVD (250 samples) | 64.6% | (Qin et al., 2023, Table 2) |
| GeoTransformer (ours) + weighted SVD (250 samples)本方法原文提出 | 86.7% | (Qin et al., 2023, Table 2) |
| GeoTransformer lite (ours, shared geometric self-attention) + weighted SVD (250 samples)本方法原文提出 | 87.5% | (Qin et al., 2023, Table 2) |
| CoFiNet + LGR (all samples) | 87.6% | (Qin et al., 2023, Table 2) |
| GeoTransformer (ours) + LGR (all samples)本方法原文提出 | 91.8% | (Qin et al., 2023, Table 2) |
| GeoTransformer lite (ours, shared geometric self-attention) + LGR (all samples)本方法原文提出 | 91.8% | (Qin et al., 2023, Table 2) |
Sun et al., 2025 · Table 5 本方法 18 筆
表格設定(擷取紀錄原文):Pairwise spatiotemporal registration on NSS; success = RRE < 10 deg and RTE < 0.2 m; methods retrained per split following original protocols. TE and RE columns (successful pairs / all pairs) not transcribed. (Sun et al., 2025, Table 5)
registration recall (RRE < 10 deg and RTE < 0.2 m),Nothing Stands Still (NSS) · Cross-Area split, all spatiotemporal pairs
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 Sun 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(Sun et al., 2025, Table 5)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| FPFH | 22.83% | (Sun et al., 2025, Table 5) |
| FCGF | 28.22% | (Sun et al., 2025, Table 5) |
| D3Feat | 31.77% | (Sun et al., 2025, Table 5) |
| Predator | 55.53% | (Sun et al., 2025, Table 5) |
| GeoTransformer本方法 | 38.13% | (Sun et al., 2025, Table 5) |
Sun et al., 2025 · Table 6 本方法 8 筆
指標registration recall
表格設定(擷取紀錄原文):Registration recall of the three best methods on NSS (original split) versus 3DMatch and 3DLoMatch; benchmark values for 3DMatch and 3DLoMatch as listed by the authors. (Sun et al., 2025, Table 6)
registration recall,3DMatch · Standard overlap (30%+)
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 Sun 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(Sun et al., 2025, Table 6)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| D3Feat | 82.2% | (Sun et al., 2025, Table 6) |
| Predator | 89% | (Sun et al., 2025, Table 6) |
| GeoTransformer本方法 | 92% | (Sun et al., 2025, Table 6) |
Qin et al., 2023 · Table 3 本方法 6 筆
資料集與序列KITTI odometry · sequences 8-10
表格設定(擷取紀錄原文):KITTI odometry sequences 8 to 10 for testing, pairs at least 10 m apart, ground truth refined with ICP; RR = share of pairs with RRE below 5 deg and RTE below 2 m; top block RANSAC-based, bottom block RANSAC-free (Qin et al., 2023, Table 3)
Relative Translation Error RTE (cm),KITTI odometry · sequences 8-10
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 Qin 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(Qin et al., 2023, Table 3)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| 3DFeat-Net | 25.9 cm | (Qin et al., 2023, Table 3) |
| FCGF | 9.5 cm | (Qin et al., 2023, Table 3) |
| D3Feat | 7.2 cm | (Qin et al., 2023, Table 3) |
| SpinNet | 9.9 cm | (Qin et al., 2023, Table 3) |
| Predator | 6.8 cm | (Qin et al., 2023, Table 3) |
| CoFiNet | 8.2 cm | (Qin et al., 2023, Table 3) |
| GeoTransformer (ours, RANSAC-50k)本方法原文提出 | 7.4 cm | (Qin et al., 2023, Table 3) |
| FMR | 66 cm有附註註記(擷取紀錄):approximate ('~66' in table) | (Qin et al., 2023, Table 3) |
| DGR | 32 cm有附註註記(擷取紀錄):approximate ('~32' in table) | (Qin et al., 2023, Table 3) |
| HRegNet | 12 cm有附註註記(擷取紀錄):approximate ('~12' in table) | (Qin et al., 2023, Table 3) |
| GeoTransformer (ours, LGR)本方法原文提出 | 6.8 cm | (Qin et al., 2023, Table 3) |
其他比較組
來源
Qin et al., 2023
(2023)GeoTransformer: Fast and Robust Point Cloud Registration With Geometric TransformerIEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9806-9821
DOI 10.1109/tpami.2023.3259038arXiv 2308.03768程式碼
同儕審查已出版已讀全文近十年
相關版本
- 會議版:Geometric Transformer for Fast and Robust Point Cloud Registration (CVPR 2022, pp. 11133-11142) 10.1109/CVPR52688.2022.01086
- 預印本:arXiv:2308.03768 (TPAMI version, 2023-07-25); CVPR version arXiv:2202.06688 per arXiv comment https://arxiv.org/abs/2308.03768
- 程式碼釋出:GeoTransformer https://github.com/qinzheng93/GeoTransformer
程式碼:https://github.com/qinzheng93/GeoTransformer(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。