FCGF
FCGF 以 Minkowski Engine 稀疏卷積構成的 ResUNet,一次計算整片點雲每個體素的 32 維幾何特徵,輸入只用座標與常數特徵,不需法向量或局部區塊(patch)前處理。作者提出最難負樣本對比損失與最難三元組損失,並以雜湊方式濾除錨點附近的假負樣本。3DMatch 上特徵匹配召回率 0.952、配準召回率平均 0.82;KITTI 上以 RANSAC 配準的成功率為 97.83% 至 98.92%。屬學習式先驗,跨感測器與場域的泛化需另行驗證。
本頁內容
Learned dense 32-D geometric features from a 3D fully-convolutional network with metric-learning losses, used for correspondence-based registration.
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 | indoor 3D scan fragments of the 3DMatch benchmark (sensor not named in the paper)、KITTI odometry LiDAR scans (model not named in the paper) |
|---|---|
| 原文測試平台 | 未記錄 |
| 狀態估計 | descriptor only; transformations estimated with RANSAC on FCGF correspondences: RANSAC with early termination for the 3DMatch registration recall (Sec. 6.5) and RANSAC for the KITTI RTE and RRE (Sec. 6.6; early termination is not stated for KITTI) |
| 資料關聯 | dense 32-D features from a ResUNet of generalized sparse convolutions (Minkowski Engine) on voxel-downsampled points with 1-vectors as input features; trained with hardest-contrastive or hardest-triplet losses using hash-based filtering of false negatives near anchors; matches found by feature similarity |
| 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 | sparse tensor (sparse voxel) representation of point clouds (Sec. 3) |
| 先驗資訊 | trained model: 3DMatch official split, or KITTI sequences 0 to 5 for training with ICP-refined GPS poses as ground truth; random scaling [0.8, 1.2] and random 3D rotation augmentation |
| 可輸出幾何 | 32-dimensional per-point features (abstract) |
| 計算需求 | 0.019 ms per feature including preprocessing; about 0.36 s per 3DMatch fragment at 2.5 cm voxel and 0.17 s at 5 cm; Intel i7-6950 10-core 3.0 GHz CPU with Nvidia Titan-X Pascal GPU; about 290x faster than PerfectMatch, 169x than 3DMatch and 42x than PPF-FoldNet |
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 | KITTI LIDAR (model not named in the paper) | 資料集感測器 | KITTI odometry | 原文未報告 | (Choy et al., 2019, Sec. 6.1, Fig. 6) |
| 運算硬體 | Intel i7-6950 (10-core, 3.0 GHz) | 執行運算平台 | 未標示 | 10-core 3.0 GHz | (Choy et al., 2019, Sec. 6.7) |
| 運算硬體 | Nvidia Titan-X Pascal | 執行運算平台 | 未標示 | 原文未報告 | (Choy et al., 2019, Sec. 6.7) |
| 運算硬體 | Intel i7 8-core 3.2 GHz CPU with Nvidia Titan-X Pascal (used by baselines [7], [6], [36]) | 執行運算平台 | 未標示 | 8-core 3.2 GHz | (Choy et al., 2019, Sec. 6.7) |
作者報告的優勢與限制
優勢
- feature-match recall 0.952 (0.953 with rotation augmentation) vs 0.947 for PerfectMatch (Table 1)
- average registration recall 0.82 vs 0.71 for PPFNet and 0.40 for FPFH on 3DMatch (Table 5)
- KITTI with 20 cm voxels: RTE 4.881 cm, RRE 0.170 deg, success 97.83% vs 3DFeat-Net 25.9 cm, 0.57 deg, 95.97% (Table 6)
- rotation invariance learned by augmentation, translation invariance inherent to sparse convolution (Sec. 6.3)
限制
- learned registration pipelines can degrade on unseen sensor patterns (Lim et al., 2025 Fig. 5, shown for Predator; not FCGF itself)
- colour input caused overfitting because the dataset was not large or diverse enough; normals gave no meaningful gain (Sec. 6.1)
- hardest-triplet loss is prone to collapse and needs random triplets mixed in (Sec. 4.2, Table 3)
- 64-D features ran out of memory at 2.5 cm voxels (Table 2)
- KITTI evaluation uses pairs at least 10 m apart and indirect RTE and RRE after RANSAC; translation error grows with voxel size (Sec. 6.2, 6.6)
- end-to-end registration left to future work (Sec. 7)
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 8 個比較組,合計 63 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 4 組列在最後,並連到性能比較頁。
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本方法 | 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 | 38.13% | (Sun et al., 2025, Table 5) |
Choy et al., 2019 · Table 6 本方法 15 筆
資料集與序列KITTI odometry (registration pairs) · test split
表格設定(擷取紀錄原文):KITTI test pairs (scans at least 10 m apart, ICP-refined GPS ground truth, 555 test pairs); RANSAC on features; success if RTE < 2 m and RRE < 5 deg; FCGF rows differ by downsampling voxel size (Choy et al., 2019, Table 6)
Relative Translation Error (RTE),KITTI odometry (registration pairs) · 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 Choy et al., 2019 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(Choy et al., 2019, Table 6)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| 3DFeat [34] | 25.9 cm | (Choy et al., 2019, Table 6) |
| FCGF 20cm本方法原文提出 | 4.881 cm | (Choy et al., 2019, Table 6) |
| FCGF 25cm本方法原文提出 | 6.066 cm | (Choy et al., 2019, Table 6) |
| FCGF 30cm本方法原文提出 | 6.466 cm | (Choy et al., 2019, Table 6) |
| FCGF 35cm本方法原文提出 | 6.978 cm | (Choy et al., 2019, Table 6) |
| FCGF 40cm本方法原文提出 | 8.025 cm | (Choy et al., 2019, Table 6) |
Qin et al., 2023 · Table 2 本方法 10 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)本方法 | 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)原文提出 | 92.3% | (Qin et al., 2023, Table 2) |
| GeoTransformer lite (ours, shared geometric self-attention) + RANSAC-50k (5000 samples)原文提出 | 92.2% | (Qin et al., 2023, Table 2) |
| FCGF + weighted SVD (250 samples)本方法 | 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)原文提出 | 86.7% | (Qin et al., 2023, Table 2) |
| GeoTransformer lite (ours, shared geometric self-attention) + weighted SVD (250 samples)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)原文提出 | 91.8% | (Qin et al., 2023, Table 2) |
| GeoTransformer lite (ours, shared geometric self-attention) + LGR (all samples)原文提出 | 91.8% | (Qin et al., 2023, Table 2) |
Choy et al., 2019 · Table 5 本方法 9 筆
指標registration recall
表格設定(擷取紀錄原文):Registration recall on the 3DMatch registration set; RANSAC with early termination; pair correct if overlap >= 30% and RMSE < 0.2 m (Choy et al., 2019, Table 5)
registration recall,3DMatch registration set · 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 Choy et al., 2019 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(Choy et al., 2019, Table 5)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| FPFH [23] | 0.36 fraction | (Choy et al., 2019, Table 5) |
| USC [29] | 0.52 fraction | (Choy et al., 2019, Table 5) |
| CGF [17] | 0.72 fraction | (Choy et al., 2019, Table 5) |
| 3DMatch [36] | 0.85 fraction | (Choy et al., 2019, Table 5) |
| PPFNet [7] | 0.9 fraction | (Choy et al., 2019, Table 5) |
| Ours (FCGF)本方法原文提出 | 0.93 fraction | (Choy et al., 2019, Table 5) |
其他比較組
來源
Choy et al., 2019
(2019)Fully Convolutional Geometric Features2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 8957-8965
DOI 10.1109/iccv.2019.00905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
- 程式碼釋出:FCGF https://github.com/chrischoy/FCGF
程式碼:https://github.com/chrischoy/FCGF(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。