Robust Reconstruction of Indoor Scenes (Redwood)
本文提出離線的 RGB-D 室內場景重建流程。先把影片切成每 50 張影格一段,以 RGB-D 里程計估計段內軌跡並以 TSDF 融合成場景片段(fragment);再對所有片段兩兩做幾何全域配準(改良的 PCL FPFH 搭配 RANSAC),得到候選迴圈閉合。作者發現即使最佳的配準演算法,精確率也低於 20%,因此在位姿圖最佳化中為每條迴圈閉合邊加入線過程(line process)變數,並以稠密表面對應距離定義對齊誤差,讓錯誤的邊在同一個最小平方問題中自動失效。修剪後再以 ICP 精修與位姿圖求得片段位姿,最後以體積整合輸出全域網格。作者同時擴充 ICL-NUIM 資料集,加入完整掃描軌跡、較真實的深度雜訊模型與辦公室場景表面真值。
本頁內容
Offline fragment-based RGB-D reconstruction: 50-frame TSDF fragments are registered pairwise by geometric global registration, and a pose graph with line-process variables on loop-closure edges and a dense surface-alignment cost prunes false loop closures before ICP refinement and volumetric integration; also introduces the augmented ICL-NUIM benchmark (full-scan trajectories, realistic noise, office ground truth).
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 video from a consumer depth camera (SUN3D real scenes; augmented ICL-NUIM synthetic sequences with a disparity-based noise and distortion model); fragment-pair registration is geometric (FPFH on fragment point sets), sensor models not named |
|---|---|
| 原文測試平台 | real indoor RGB-D sequences (SUN3D scenes and an apartment sequence; capture platform not described in the paper)、simulation (augmented ICL-NUIM living room and office; trajectories model thorough handheld scanning) |
| 狀態估計 | Offline pipeline: RGB-D odometry (Kerl et al. 2013) inside 50-frame fragments; pose graph over fragment poses with odometry edges and loop-closure edges weighted by line-process variables l in [0, 1] with prior (sqrt(l) - 1)^2 and balance mu = tau^2 kappa, solved with g2o; edges with l < 0.25 pruned; ICP refinement of the remaining edges and final pose-graph optimization; optional SLAC non-rigid refinement |
| 資料關聯 | Pairwise geometric registration of every fragment pair with a modified PCL FPFH plus RANSAC algorithm (four-point samples, normal-angle, edge-length and overlap checks); candidates kept when more than 30% overlap; alignment cost uses dense point correspondences within 5 cm |
| 時間表示 | discrete poses (fragment poses; frame poses within fragments from odometry) |
| 去畸變 | 不適用 (RGB-D input) |
| 迴圈閉合 | Geometric: every fragment pair tested by global registration, so loops are found even when images are not similar; false candidates removed by the line-process optimization |
| 全域最佳化 | Offline robust pose-graph optimization of fragments (line processes, g2o), then ICP refinement and pose-graph optimization; optional non-rigid refinement |
| 地圖表示 | fragment meshes from volumetric TSDF integration (Curless and Levoy) of 50-frame segments; final global volumetric integration |
| 先驗資訊 | none |
| 可輸出幾何 | global surface mesh (for example 15.8 million triangles for the 17,391-frame apartment) |
| 計算需求 | Offline on a workstation with Intel Core i7-3770 3.5 GHz CPU and 16 GB RAM; total 29 to 187 minutes per evaluated sequence and 387 minutes for the apartment (supplementary Table 1) |
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 | workstation with Intel Core i7-3770 3.5GHz CPU and 16GB of RAM | 執行運算平台 | 未標示 | all pipeline steps; registration timings single-threaded | (Choi et al., 2015, Table 1 caption; supplementary Table 1) |
作者報告的優勢與限制
優勢
- Loop-closure precision rises from 19.6% to 97.7% while recall drops only from 59.2% to 57.8% on augmented ICL-NUIM (Table 2)
- Mean distance to the ground-truth surface 0.03 to 0.07 m (average 0.05 m), about half that of the next best pipeline, SUN3D SfM (Table 4)
- Finds loop closures from geometry even when the corresponding images are not similar (Fig. 2)
- Ranked above all automatic pipelines on eight SUN3D scenes by crowd-sourced comparisons, and above the manually assisted models on 6 of 8 (Table 7)
限制
- Offline: 29 to 187 minutes per evaluated sequence and 387 minutes for the 17,391-frame apartment on a desktop CPU (supplementary Table 1)
- If the input contains no loop closures, odometry drift accumulates and distorts the model (Sec. 7)
- Catastrophic odometry failure inside a fragment is not handled (Sec. 7)
- Real-world scenes are evaluated only perceptually by crowd-sourced pairwise comparisons, not against measured geometry (Sec. 6.3)
營建工程相關證據
論文未在施工現場測試,真實場景評估為 SUN3D 的住宅、旅館與校園室內,且以群眾外包成對比較評分,並非對照量測幾何真值。以片段為單位的離線重建與穩健迴圈閉合,適合工地資料事後處理的整層室內重建;Fig. 1 的公寓重建軌跡長 151.6 公尺,但單一序列需數小時處理。corpus 中 Dai et al., 2017a 以 Redwood 作為離線基準。
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 7 個比較組,合計 53 筆紀錄。以下列出本方法紀錄最多的 4 組,其餘 3 組列在最後,並連到性能比較頁。
Choi et al., 2015 · Table 2 本方法 16 筆
表格設定(擷取紀錄原文):Loop-closure recall of fragment pairs (ground-truth loop = more than 30% overlap; true positive if ground-truth correspondence RMSE below 0.2 m) (Choi et al., 2015, Table 2)
Recall (%),augmented ICL-NUIM (synthetic, realistic noise, full-scan trajectories) · Living room 1
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 Choi et al., 2015 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(Choi et al., 2015, Table 2)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Geometric registration candidates before pruning本方法 | 61.2% | (Choi et al., 2015, Table 2) |
| After line-process pruning (Ours)本方法原文提出 | 57.6% | (Choi et al., 2015, Table 2) |
Choi et al., 2015 · Supp. Table 1 本方法 13 筆
指標Total time (minutes)
表格設定(擷取紀錄原文):Total running time of all pipeline steps (fragment creation, geometric registration, robust optimization, ICP refinement, integration) per sequence (Choi et al., 2015, Supp. Table 1)
Total time (minutes),authors' apartment sequence · Apartment (17,391 frames)
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 Choi et al., 2015 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 |
|---|---|---|
| Ours本方法原文提出硬體:Intel Core i7-3770 3.5 GHz, 16 GB RAM | 387 min | (Choi et al., 2015, Supplementary Table 1) |
Choi et al., 2015 · Table 7 本方法 8 筆
指標BRE score
表格設定(擷取紀錄原文):Perceptual evaluation on real SUN3D scenes: Balanced Rank Estimation scores from 17,640 crowd-sourced pairwise comparisons (Amazon Mechanical Turk), range -1 to 1, higher is better; not a geometric measurement (Choi et al., 2015, Table 7)
BRE score,SUN3D · hotel umd
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 Choi et al., 2015 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(Choi et al., 2015, Table 7)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| DVO SLAM [34] | -0.61 | (Choi et al., 2015, Table 7) |
| Kintinuous [61] | -0.45 | (Choi et al., 2015, Table 7) |
| SUN3D SfM [65] | -0.02 | (Choi et al., 2015, Table 7) |
| Ours本方法原文提出 | 0.66 | (Choi et al., 2015, Table 7) |
| SUN3D manual (manually assisted reconstructions) | 0.56 | (Choi et al., 2015, Table 7) |
Choi et al., 2015 · Supp. Table 3 本方法 4 筆
指標median distance to ground-truth surface (m)
表格設定(擷取紀錄原文):Median distance of each reconstructed model to the ground-truth surface (supplementary Appendix E) (Choi et al., 2015, Supp. Table 3)
median distance to ground-truth surface (m),augmented ICL-NUIM (synthetic, realistic noise, full-scan trajectories) · Living room 1
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 Choi et al., 2015 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(Choi et al., 2015, Supp. Table 3)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| Kintinuous [61] | 0.17 m | (Choi et al., 2015, Supplementary Table 3) |
| DVO SLAM [34] | 0.16 m | (Choi et al., 2015, Supplementary Table 3) |
| SUN3D SfM [65] | 0.08 m | (Choi et al., 2015, Supplementary Table 3) |
| Ours本方法原文提出 | 0.03 m | (Choi et al., 2015, Supplementary Table 3) |
| GT trajectory (input depth fused along ground truth, reference) | 0.03 m | (Choi et al., 2015, Supplementary Table 3) |
其他比較組
來源
Choi et al., 2015
(2015)Robust reconstruction of indoor scenes2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5556-5565
DOI 10.1109/cvpr.2015.7299195程式碼
同儕審查已出版已讀全文經典查證後修正
相關版本
- 程式碼釋出:qianyizh/ElasticReconstruction https://github.com/qianyizh/ElasticReconstruction
- 資料集:Augmented ICL-NUIM dataset and indoor reconstruction pages (redwood-data.org/indoor) http://redwood-data.org/indoor/
程式碼:https://github.com/qianyizh/ElasticReconstruction(授權:MIT-style licence (LICENSE.txt checked, Copyright 2013 Stanford University))。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。