← 回文獻卡索引

面向真實場景的三維高斯潑濺方法3D Gaussian Splatting Methods for Real-World Scenarios

Ivana Petrovska, Boris Jutzi｜ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences X-G-2025: 641-648｜DOI: 10.5194/isprs-annals-X-G-2025-641-2025

狀態：AI_DRAFT_FROM_REVIEW｜分級：A｜閱讀深度：FULL_TEXT_CHECKED｜Jacky 審核：False

森林數位孿生產業應用

讀學習筆記 →

3D reconstruction3D Gaussian Splattingocclusionvegetationpoint cloud comparison3DGSNeRFMulti-View StereoSfMcloud-to-meshGermanyindooroutdoor

專討核心文獻定位

[83] Ch7 · 數位孿生 ★ 新增
Petrovska & Jutzi · 2025
以室內與植被遮擋兩種真實場景比較四種 3DGS 方法、MVS 與 NeRF，原始 3DGS 精度第二僅次 MVS、3DGS-MCMC 完整度最高，並證實 3DGS 能重建植被遮擋後的幾何，指向森林冠層與生物量應用潛力

使用警示

本頁是文獻知識庫卡片，不等於可直接引用的最終查核稿。只有狀態升級為 CITABLE 後，才可直接進入論文引用候選。

為什麼納入這篇

This paper is an application-layer benchmark of 3D Gaussian Splatting methods against Multi-View Stereo and NeRF in real-world scenarios, with one scenario placing the object behind vegetation occlusion. It quantifies the accuracy and completeness trade-offs of the front-end 3D reconstruction layer that a forest digital twin depends on, and explicitly argues 3DGS potential for canopy reconstruction and biomass estimation behind vegetation occlusion.

結構式摘要｜中英文對照

研究問題
在真實的室內無遮擋與戶外植被遮擋兩種場景中，不同平台上的三維高斯潑濺方法能否在幾何精度與完整度上挑戰傳統與最新的三維重建方法 MVS 與 NeRF，並可靠重建植被遮擋後方的幾何？
In two real-world scenarios, one occlusion-free indoor and one outdoor with vegetation occlusion, can 3D Gaussian Splatting methods on different radiance-field platforms challenge MVS and NeRF in geometric accuracy and completeness, and reliably reconstruct the geometry behind vegetation occlusion?
資料來源
原文 Section 3.1 確認，資料取自 STELLA 資料集的兩個真實場景，一為無遮擋的 Original、一為以植物作為非透明遮擋的 Vegetation。待評估物件是一尊 0.7 公尺高的佛像，放在矩形底板上。每個場景沿環形軌跡拍攝 125 張高解析影像，因記憶體與效率考量降採樣為 1840x1228 像素並轉為無損 png。地面真值以結構光成像取得、精度 0.1 mm 的網格。
Section 3.1 confirms two real-world scenarios from the STELLA dataset, Original which is occlusion-free and Vegetation with plants as non-transparent occlusion. The object to be evaluated is a 0.7 m tall Buddha statue on a rectangular plate. Each scenario has 125 high-resolution images captured on a circular trajectory, downsampled to 1840x1228 pixels and converted to lossless png. Ground truth is a mesh generated by Structured Light Imaging at 0.1 mm accuracy.
方法
原文 Section 2 與 3 確認，共比較六種三維重建方法：傳統 MVS（COLMAP 的 SfM 加密集 MVS）、NeRF 代表 Nerfacto（Nerfstudio v1.1.0），以及四種 3DGS 方法，分別是原始實作 3DGS-Basic、gsplat 為核心的 Splatfacto、以馬可夫鏈蒙地卡羅取代啟發式增稠的 3DGS-MCMC（在 Postshot 中），以及雲端專有實作 3DGS-LumaAI。各方法皆以 SfM 估相機位姿與稀疏點雲，其中 3DGS-Basic、Splatfacto、3DGS-MCMC 另以稀疏點雲初始化。評估方式為先以 ICP 將各點雲對齊地面真值網格，再以 cloud-to-mesh 計算精度（Mean、SD、RMSE）與完整度（5 mm 門檻內覆蓋比例），完整度以網格次取樣 1000 萬點為基準。實驗硬體為 Intel i9 CPU、32GB RAM、Nvidia RTX 3090。
Sections 2 and 3 confirm six reconstruction methods: traditional MVS (COLMAP SfM plus dense MVS), the NeRF representative Nerfacto (Nerfstudio v1.1.0), and four 3DGS methods: the original 3DGS-Basic, gsplat-based Splatfacto, 3DGS-MCMC which replaces heuristic densification with Markov Chain Monte Carlo (in Postshot), and the proprietary cloud-based 3DGS-LumaAI. All methods estimate camera poses and a sparse point cloud via SfM; 3DGS-Basic, Splatfacto, and 3DGS-MCMC additionally use the sparse cloud for initialization. Evaluation aligns each cloud to the ground-truth mesh with ICP, then computes cloud-to-mesh accuracy (Mean, SD, RMSE) and completeness (coverage within a 5 mm threshold), with completeness referenced against 10M points subsampled on the mesh. Hardware was an Intel i9 CPU, 32GB RAM, Nvidia RTX 3090.
主要結果
原文 Section 4 與 Table 1 確認三組結論。精度上，MVS 最準（Original RMSE 1.43 mm、Vegetation 3.23 mm），原始 3DGS-Basic 全場第二、優於 NeRF（Original 4.74 mm、Vegetation 7.49 mm），是最準的 3DGS 方法；NeRF 精度差（5.36、12.67 mm），Splatfacto 與 3DGS-MCMC 在 Vegetation 誤差大（13.80、17.36 mm）。完整度上，3DGS-MCMC 在 Original 達 99.45%、為全場最高，Vegetation 65.93%、為 3DGS 法中第三；所有 3DGS 方法在 Vegetation 完整度都明顯下降（3DGS-Basic 由 82.75% 掉到 42.71%），對遮擋敏感。效率上（Table 2），NeRF 最快僅 15 分鐘、Splatfacto 25 分鐘次之，MVS 最慢約 1 小時 15 分。最重要的是原文證實 3DGS 能可靠重建植被遮擋後方的幾何，指向冠層重建、生物量估算與農業監測的大尺度林業應用潛力。
Section 4 and Table 1 confirm three sets of results. On accuracy, MVS is best (Original RMSE 1.43 mm, Vegetation 3.23 mm) and the original 3DGS-Basic is second overall, outperforming NeRF (Original 4.74 mm, Vegetation 7.49 mm), making it the most accurate 3DGS method; NeRF is poor (5.36, 12.67 mm), and Splatfacto and 3DGS-MCMC have large Vegetation errors (13.80, 17.36 mm). On completeness, 3DGS-MCMC reaches 99.45% in Original, the highest overall, and 65.93% in Vegetation, third among 3DGS methods; all 3DGS methods drop sharply in Vegetation (3DGS-Basic falls from 82.75% to 42.71%), showing sensitivity to occlusion. On efficiency (Table 2), NeRF is fastest at 15 min, Splatfacto next at 25 min, and MVS slowest at about 1 h 15 min. Most importantly, the paper confirms 3DGS can reliably reconstruct geometry behind vegetation occlusion, indicating potential for large-scale forestry applications such as canopy reconstruction, biomass estimation, and agricultural monitoring.
限制
原文 Section 4 與 5 指出多項限制。NeRF 與所有 3DGS 方法精度都低於 MVS，因幾何是靠最小化影像重建損失學得，點雲未緊貼物件表面、物件內部還有偽影點，會拉低精度且不貢獻完整度。所有 3DGS 方法在植被遮擋下完整度大幅下滑，限制其在遮擋場景的穩健性。3DGS-MCMC 與 3DGS-LumaAI 程式碼非開源、參數調整受限，無法深入分析；3DGS-MCMC 的最大高斯數沿用預設 3M，對較大場景可能不足。3DGS-LumaAI 在 Vegetation 完整度僅 27.02%、最低，且無法確認其初始化與評估細節。原文結論也指出處理大尺度且遮擋密集的森林區域需要高記憶體與運算量。研究僅以單一佛像物件與兩個小型場景驗證，跨物件與大尺度林分泛化未證。
Sections 4 and 5 note several limitations. NeRF and all 3DGS methods have lower accuracy than MVS because geometry is learned by minimizing image reconstruction loss; points do not align tightly with the object surface and artifact points inside the object lower accuracy without contributing to completeness. All 3DGS methods drop sharply in completeness under vegetation occlusion, limiting robustness in occluded scenes. 3DGS-MCMC and 3DGS-LumaAI are not open-source with limited parameter adjustment, restricting deeper analysis; 3DGS-MCMC used a default 3M maximum Gaussian count that may be insufficient for larger scenes. 3DGS-LumaAI has the lowest completeness at 27.02% in Vegetation and its initialization and evaluation cannot be confirmed. The conclusion notes that processing large-scale, densely occluded forest regions requires high memory and computation. The study validated only a single Buddha-statue object across two small scenarios, so cross-object and large-scale stand generalization is unproven.

Key Findings

發現證據確定性
The original 3DGS implementation achieves second-best accuracy overall, just behind MVS and outperforming NeRF in both scenarios, making it the most accurate 3DGS method.Section 4, Table 1: 3DGS-Basic RMSE 4.74 mm (Original) and 7.49 mm (Vegetation) vs NeRF 5.36 and 12.67 mm; MVS best at 1.43 and 3.23 mm.checked_against_full_text
3DGS-MCMC reaches the highest completeness among all methods in the occlusion-free scenario, competitive with MVS and NeRF.Section 4, Table 1: 3DGS-MCMC completeness 99.45% in Original (highest overall) and 65.93% in Vegetation (third best among 3DGS), but highest RMSE 17.36 mm in Vegetation.checked_against_full_text
3DGS can reliably reconstruct geometry behind vegetation occlusion, but all 3DGS methods lose substantial completeness under occlusion, limiting robustness.Abstract and Conclusion (Section 6): 3DGS-Basic completeness falls from 82.75% (Original) to 42.71% (Vegetation); authors argue potential for canopy reconstruction, biomass estimation, and agricultural monitoring.checked_against_full_text

Key Figures and Tables

公開網站原則：未確認授權前，不直接複製原文圖表；優先使用自製圖表導讀或重繪圖。

項目內容關鍵數字Jacky 判讀重用策略
Table 1Mean, SD, RMSE (accuracy in mm) and point count plus completeness percentage for MVS, NeRF, and the four 3DGS methods across Original and Vegetation scenarios.RMSE Original/Vegetation (mm): MVS 1.43/3.23, NeRF 5.36/12.67, 3DGS-Basic 4.74/7.49, Splatfacto 7.28/13.80, 3DGS-MCMC 5.80/17.36, 3DGS-LumaAI 7.02/7.58. Completeness Original/Vegetation (%): MVS 97.35/75.95, NeRF 98.23/69.95, 3DGS-Basic 82.75/42.71, Splatfacto 79.83/31.80, 3DGS-MCMC 99.45/65.93, 3DGS-LumaAI 88.77/27.02.這張表是精度歸 MVS 與 3DGS-Basic、完整度歸 3DGS-MCMC 這個核心結論的量化來源，也清楚顯示所有 3DGS 法在植被遮擋下完整度幾乎腰斬，可用來談 3DGS 進入森林前端重建的邊界條件。以自繪長條圖重述 RMSE 與完整度數字，標註出處 Petrovska 與 Jutzi 2025 Table 1，不直接貼原表。
Table 2Open-source availability and reconstruction or training time for each of the six methods.Training time: MVS 1h 15min（最慢）, NeRF 15min（最快）, 3DGS-Basic 49min, Splatfacto 25min, 3DGS-MCMC 43min, 3DGS-LumaAI 45min。開源者為 MVS、NeRF、3DGS-Basic、Splatfacto 四種；3DGS-MCMC（Postshot）與 3DGS-LumaAI 非開源。效率與開源性這張表支撐工程落地選型討論，NeRF 最快、Splatfacto 是最快的 3DGS 法，但非開源方法限制可調參與可複製性。以文字或自繪圖重述時間與開源狀態，標註出處，不複製原表。
Figure 3, Figure 4Figure 3 為六種方法在 Original 與 Vegetation 兩場景的點雲幾何重建；Figure 4 為對地面真值網格的 cloud-to-mesh 帶符號距離（色階 90mm 到 -90mm），正值在網格外、負值在物件內。原文指出除 MVS 外，重建點雲皆在物件內部出現偽影點而扭曲精度；Vegetation 因遮擋，所有點雲點數都比無遮擋的 Original 少。這兩張圖視覺化說明為何 NeRF 與 3DGS 精度輸給 MVS，根因是物件內部偽影點，可用來解釋稠密不等於準確。以自繪示意圖呈現帶符號距離與內部偽影概念，標註出處，不複製原圖。

Extracted Evidence Table

可支撐主張指標或結果原文位置可引用備註
The original 3DGS achieves second-best accuracy overall and is the most accurate 3DGS method, beating NeRF.3DGS-Basic RMSE 4.74 mm (Original) / 7.49 mm (Vegetation) vs NeRF 5.36 / 12.67 mm; MVS best 1.43 / 3.23 mm (Table 1).Section 4, Table 1, Conclusion (Section 6).True可引為 3DGS 在前端三維重建精度上具競爭力的實證；勿宣稱其精度已超越 MVS。
All 3DGS methods lose substantial completeness under vegetation occlusion, though 3DGS can still recover occluded geometry.3DGS-Basic completeness 82.75% → 42.71%; 3DGS-LumaAI lowest at 27.02% in Vegetation (Table 1).Section 4, Table 1, Abstract, Conclusion.True用以同時論述 3DGS 看穿遮擋的潛力與遮擋下完整度大幅下降的限制。
The authors argue 3DGS shows potential for large-scale forestry applications behind vegetation occlusion.Demonstrated reliable reconstruction of geometry behind vegetation occlusion, indicating canopy reconstruction, biomass estimation, and agricultural monitoring potential.Abstract; Conclusion (Section 6).True此為作者前瞻論述而非已驗證的森林尺度部署，引用時須標明屬潛力與展望。

Critical Appraisal

Strengths

Weaknesses

Validation qualityquantitatively validated by cloud-to-mesh comparison (Mean, SD, RMSE for accuracy and a 5 mm-threshold coverage ratio for completeness) against a 0.1 mm Structured Light Imaging ground-truth mesh across two real-world scenarios and six reconstruction methods
Transferability to Taiwanmedium：以 3DGS 重建植被遮擋後幾何的方法概念可移植到台灣森林冠層與行道樹重建，但本文僅在物件尺度驗證，需就台灣常綠闊葉與針葉林、以及大尺度天然林重新驗證記憶體與遮擋穩健性。
Risk of overclaiming勿將本文描述為完整森林數位孿生或已驗證的森林尺度部署；它是物件尺度的 3DGS 與 MVS、NeRF 重建比較基準，森林應用屬作者的潛力展望，且結論明指遮擋下完整度大幅下降、大尺度遮擋森林需高記憶體與運算量。

與 Jacky 博論 / Review 的用途

博士論文支撐博論在 FDT 前端三維重建層的技術選型討論，量化說明 3DGS 各變體與 MVS、NeRF 在精度、完整度與效率上的取捨，特別是看穿植被遮擋的能力，是後續接上即時感測與碳估算前的基礎環節。
TJFS Review在 TJFS review 的 Ch7 數位孿生章節，作為以 3DGS 做真實場景重建並專門檢驗植被遮擋的應用層實證，與 Tian 2025 的林分 NVS 比較、Chen 2024 的城市樹木建模、Buonocore 等的概念框架對照，補強框架到落地之間的前端重建技術細節。
可引用句候選2025 年，Petrovska 等人發表的文獻中指出，在室內無遮擋與戶外植被遮擋兩種真實場景的三維重建中，原始 3D 高斯潑濺實作的幾何精度僅次於多視角立體視覺、優於神經輻射場（植被場景 RMSE 7.49 mm 對 NeRF 12.67 mm），且 3D 高斯潑濺能可靠重建植被遮擋後方的幾何，顯示其在森林冠層重建與生物量估算的應用潛力。
不可用來主張勿用本文單獨作為「完整森林數位孿生已落地」或「3DGS 已驗證於大尺度森林」的證據，也勿宣稱 3DGS 精度已超越 MVS 或其遮擋下完整度已足以可靠重建森林冠層。

授權與圖表重用

Article licenseCC-BY-4.0
Figure reuse policyDO_NOT_REUSE_ORIGINAL_FIGURES_PUBLICLY_UNTIL_LICENSE_CHECKED
Notes原文每頁頁尾版權聲明確認為 Author(s) 2025、CC BY 4.0 License（ISPRS Annals, Copernicus Publications）。CC BY 允許在標註出處下重製與改作；惟仍依本知識庫保守政策，公開呈現以自繪示意圖替代，正式投稿前再核對授權與出處標註。

待查核清單