← 回文獻卡索引

以 OB-NeRF 重建複雜植株的三維重建平台A 3D reconstruction platform for complex plants using OB-NeRF

Sixiao Wu, Changhao Hu, Boyuan Tian, Yuan Huang, Shuo Yang, Shanjun Li, Shengyong Xu｜Frontiers in Plant Science 16: 1449626｜DOI: 10.3389/fpls.2025.1449626

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 reconstructionplant phenotypingdigital twinindividual planthigh-throughput phenotypingNeRFOB-NeRFSfMmulti-resolution hash encodingmarching cubesChinaWuhan

專討核心文獻定位

[95] Ch7 · 數位孿生 ★ 新增
Wu et al. · 2025
提出 Object-Based NeRF，以相機影像軌跡作先驗自動校正位姿，無需去背即可在 30 秒內完成複雜植株高保真三維重建，為農業數位孿生提供前端資料

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 source for the neural 3D reconstruction front-end that a digital twin depends on. It directly frames high-fidelity per-plant 3D reconstruction as the data foundation for agricultural digital twins and shows how an Object-Based NeRF variant removes the main NeRF bottlenecks of long runtime, background segmentation, and lighting sensitivity at the single-plant scale.

結構式摘要｜中英文對照

研究問題
在低成本且無需人工去背或受控光照的條件下，如何讓 NeRF 類神經重建在複雜背景與不均光照下，於極短時間內完成單株複雜植株的高精度三維重建，以支撐農業數位孿生與高通量表型分析？
Under low cost and without manual background removal or controlled lighting, how can a NeRF-style neural reconstruction achieve high-precision 3D reconstruction of a single complex plant in a very short time under complex backgrounds and uneven illumination, so as to support agricultural digital twins and high-throughput phenotyping?
資料來源
原文 Section 2 確認，材料為華中農業大學數位果園示範基地的柑橘苗（椪柑 Citrus reticulata「Yura」），共 20 株果樹苗，依高度分為小型 60 至 90 公分 7 株、中型 90 至 130 公分 7 株、大型 130 至 170 公分 6 株，苗齡 6 個月至 2 年、平均 18 個月、平均高約 120 公分。以「camera to plant」自動視訊擷取系統取像：金屬環半徑 75 公分、上下兩台相機，實驗用兩支 iPhone 14 Pro，影像解析度 1080×1920、30 fps；每次擷取約 15 秒完成 360 度環拍，雙目視訊等距下採樣後各取 50 張、共 100 張，去除模糊後保留約 90 張作為重建輸入。
Section 2 confirms the materials are citrus saplings (Citrus reticulata 'Yura') from the digital orchard demonstration base at Huazhong Agricultural University, 20 fruit-tree saplings grouped by height into small 60-90 cm (7), medium 90-130 cm (7), and large 130-170 cm (6), aged 6 months to 2 years (mean 18 months, mean height about 120 cm). Images were captured with a camera-to-plant automated video acquisition system: a metal ring of radius 75 cm with upper and lower cameras, using two iPhone 14 Pro phones at 1080x1920 resolution and 30 fps; each capture completes a 360-degree orbit in about 15 seconds, and the binocular videos are equidistantly downsampled to 50 images per view (100 total), then reduced to about 90 clear frames as reconstruction input after blur removal.
方法
原文 Section 3 確認，管線分三步：多視角影像擷取、相機參數估計與校正、以 OB-NeRF 進行三維重建。先以張正友標定法取得相機內參，再以 SfM 估計相機位姿；接著提出基於軟硬體協同設計的相機位姿全域校正策略，把預設成像軌跡的法向量作為方向參考、軌跡直徑作為尺度參考，用最小二乘與 Rodrigues 旋轉公式還原模型真實位姿與尺寸，免去額外標定器。OB-NeRF 在 NeRF 基礎上加入四項改良：以 AABB 界定 ROI 的新光線取樣策略（聚焦目標植株、跳過空白空間，無需去背）、淺層 MLP 加速訓練、引自 Instant-NGP 的多解析度雜湊編碼（最末層細分為 2048 個雜湊體素以強化細尺度表現）、以及把每張影像曝光率設為可學習參數的曝光調整模組以因應不均光照；訓練後以 Marching Cubes 從輻射場萃取 Mesh。實驗硬體為 Intel Core i7-12700H、16 GB 記憶體、NVIDIA RTX 3070Ti Laptop GPU、Windows 11。
Section 3 confirms a three-stage pipeline: multi-view image acquisition, camera parameter estimation and calibration, and 3D reconstruction with OB-NeRF. Camera intrinsics are obtained by Zhang Zhengyou's calibration method and camera poses are estimated by SfM; then a hardware-software co-design global pose calibration strategy uses the normal vector of the predetermined imaging trajectory as a direction reference and the trajectory diameter as a scale reference, applying least squares and Rodrigues' rotation formula to restore the model's true pose and size without an extra calibrator. OB-NeRF adds four improvements over NeRF: a new ROI ray-sampling strategy bounded by an AABB (focusing on the target plant and skipping empty space, so no background segmentation is needed), a shallow MLP to speed training, multi-resolution hash encoding from Instant-NGP (the finest level subdivided into 2048 hash voxels for finer-scale fidelity), and an exposure adjustment module treating each image's exposure rate as a learnable parameter for uneven lighting; after training, Marching Cubes extracts the mesh from the radiance field. Experiments ran on an Intel Core i7-12700H, 16 GB memory, NVIDIA RTX 3070Ti Laptop GPU, Windows 11.
主要結果
原文 Section 4 與 Section 5 確認多項結果。速度上，OB-NeRF 將單株植株重建時間從 NeRF 的逾 10 小時降至 30 秒，整條管線（含 SfM 約 210 秒）約 250 秒完成，資料擷取約 15 秒。品質上，目標植株可在背景嚴重失真（PSNR 低於 20 dB）時仍維持高保真（PSNR = 29.95 dB），器官層級平均重建誤差小於 2 公釐、紋理解析度達 0.5 公釐／像素、空間解析度 0.0019 公釐。精度上，樹高、葉長、葉寬的 R² 分別為 0.9933、0.9881、0.9883，MAE 分別為 2.0947、0.1898、0.1199 公分，達公釐級精度。對照上（Table 2），Based-NeRF 訓練需 9 至 12 小時、平均 PSNR 24.71 dB，OB-NeRF 為 30 秒、29.95 dB，優於 Mip-NeRF、Instant-NGP、NeuS、NeRFacto；與 COLMAP、Kinect 相比，OB-NeRF 重建時間較 COLMAP 減少 96.08%、較 Kinect 減少 49.73%，模型點數較 COLMAP 多 159.26%、較 Kinect 多 393.72%，空間解析度 0.0019 公釐優於 COLMAP 的 0.0100 與 Kinect 的 0.0037 公釐。
Sections 4 and 5 confirm several results. On speed, OB-NeRF cuts single-plant reconstruction time from NeRF's over 10 hours to 30 seconds, with the full pipeline (including about 210 s of SfM) finishing in about 250 seconds and data acquisition about 15 seconds. On quality, the target plant stays high fidelity (PSNR = 29.95 dB) even when the background is severely distorted (PSNR below 20 dB), with organ-level average reconstruction error below 2 mm, texture resolution 0.5 mm per pixel, and spatial resolution 0.0019 mm. On accuracy, R2 for tree height, leaf length, and leaf width is 0.9933, 0.9881, and 0.9883 respectively, with MAE 2.0947, 0.1898, and 0.1199 cm, reaching millimeter-level precision. On comparison (Table 2), Based-NeRF needs 9-12 hours training at mean PSNR 24.71 dB while OB-NeRF is 30 s at 29.95 dB, outperforming Mip-NeRF, Instant-NGP, NeuS, and NeRFacto; versus COLMAP and Kinect, OB-NeRF reduces reconstruction time by 96.08% over COLMAP and 49.73% over Kinect, yields 159.26% more model points than COLMAP and 393.72% more than Kinect, and reaches spatial resolution 0.0019 mm versus COLMAP's 0.0100 and Kinect's 0.0037 mm.
限制
原文 Section 5 指出，本方法主要針對單株植株的三維重建，延伸到大尺度田間場景時實驗顯示性能受限，未來需引入新的最佳化策略或更先進演算法（如 3DGS）。SfM 仍是上游瓶頸，特徵匹配時間隨影像數量呈二次成長，須將影像數控制在約 90 張以平衡品質與耗時。作者規劃將重建流程上雲、並以軟硬體協同的精密機械與馬達控制在已知位姿下取像以取代 SfM。整體驗證僅在單一柑橘苗樹種、室內基地、特定相機與環拍硬體完成，跨樹種與森林尺度的泛化尚未證實。
Section 5 notes that the method is primarily designed for single-plant 3D reconstruction and experiments show performance limitations when extended to large-scale field scenes, so future work should introduce new optimization strategies or more advanced algorithms such as 3DGS. SfM remains the upstream bottleneck, with feature-matching time growing quadratically with image count, requiring the image number to be kept around 90 to balance quality and time. The authors plan to move reconstruction to the cloud and use hardware-software co-designed precise mechanics and motor control to capture images under known poses, replacing SfM. Overall validation was done only on a single citrus species, an indoor base, and specific camera and orbital hardware, so generalization across species and to forest scale is unproven.

Key Findings

發現證據確定性
OB-NeRF reconstructs a single complex plant in 30 seconds versus over 10 hours for baseline NeRF, with the full pipeline finishing in about 250 seconds.Abstract and Sections 4-5, Table 2: reconstruction time reduced from over 10 h to 30 s; Based-NeRF needs 9-12 h; full pipeline about 250 s when processing 90 images.checked_against_full_text
A ROI-based ray sampling strategy lets the target plant be reconstructed at high fidelity without background segmentation, even when the background itself is badly distorted.Section 4.1: target plant PSNR = 29.95 dB while background PSNR is below 20 dB; OB-NeRF embeds the AABB ROI as prior knowledge to skip empty space and avoid masks.checked_against_full_text
Mesh-extracted phenotypes reach millimeter-level accuracy against manual measurement, supporting agricultural digital twins.Section 4.1, Fig. 9: R2 = 0.9933 / 0.9881 / 0.9883 and MAE 2.0947 / 0.1898 / 0.1199 cm for tree height, leaf length, and leaf width; organ-level error below 2 mm; spatial resolution 0.0019 mm.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 2Average training time and average PSNR for six reconstruction methods on the same plant data.Based-NeRF 10 h / 24.71 dB；Mip-NeRF 3 h / 25.81 dB；Instant-NGP 30 s / 26.79 dB；NeuS 30 min / 27.31 dB；NeRFacto 5 min / 27.64 dB；Ours (OB-NeRF) 30 s / 29.95 dB。這張表是 OB-NeRF 同時兼顧速度與品質的最硬證據，可用來論述把神經式重建放進數位孿生前端的可行性，但須提醒這是單株室內柑橘苗的對照，非森林尺度。以自繪比較表或長條圖重述時間與 PSNR 數字，標註出處 Wu et al. 2025 Table 2，不直接貼原表。
Fig. 9Fig. 9(a)(b)(c) 分別為樹高、葉長、葉寬之模型估計值對人工量測真值的線性擬合。樹高 R2 = 0.9933、MAE 2.0947 cm；葉長 R2 = 0.9881、MAE 0.1899 cm；葉寬 R2 = 0.9883、MAE 0.1199 cm。這張圖是公釐級表型精度結論的量化來源，可用來談 NeRF 類重建在單株尺度已足以支撐表型與孿生資料層。以自繪散點或趨勢示意圖呈現 R2 與 MAE，標註出處，不複製原圖。
Figure 5, Figure 6Figure 5 為 OB-NeRF 整體結構，Figure 6 拆解三項改良：(a) 光線取樣策略、(b) 密度場與顏色場 MLP 網路、(c) 多解析度雜湊編碼。雜湊表 16 層、最大條目 2^19、每條目特徵維度 2、最粗解析度 16、最細解析度 2048（Table 1）。這兩張圖是理解 OB-NeRF 為何能又快又準的機制圖，適合在 review 中說明 NeRF 加速與細節保真的技術路徑。以自繪流程示意圖重述四項改良與雜湊參數，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
OB-NeRF dramatically accelerates per-plant NeRF reconstruction while improving PSNR.30 s vs over 10 h for Based-NeRF; PSNR 29.95 dB vs 24.71 dB (Table 2); full pipeline about 250 s.Abstract, Section 4.3, Table 2, Section 5.True可引為 NeRF 前端加速的實證；勿宣稱其已整合為完整數位孿生系統或已驗證於森林尺度。
Reconstruction reaches millimeter-level phenotype accuracy without background masks.R2 0.9933 / 0.9881 / 0.9883 與 MAE 2.0947 / 0.1898 / 0.1199 cm；目標植株 PSNR 29.95 dB 而背景低於 20 dB。Section 4.1, Fig. 9.True用以說明無需去背的高精度重建；限定為單株柑橘苗、室內基地。
OB-NeRF outperforms COLMAP and Kinect-based reconstruction in time and detail.重建時間較 COLMAP 減 96.08%、較 Kinect 減 49.73%；點數較 COLMAP 多 159.26%、較 Kinect 多 393.72%；空間解析度 0.0019 mm 對 0.0100 與 0.0037 mm。Section 4.3, Fig. 14, Table 5.True可引為與傳統攝影測量與深度相機的對照；勿外推到大尺度田間或森林。

Critical Appraisal

Strengths

Weaknesses

Validation qualityquantitatively validated against manual measurement using R2 and MAE for tree height, leaf length, and leaf width across 20 saplings in three size groups, plus PSNR and spatial/texture resolution comparison with NeRF variants, COLMAP, and Kinect.
Transferability to Taiwanmedium：無需去背、消費級相機加環拍硬體的低成本 NeRF 重建管線，方法上可移植到台灣苗圃、果園或都市行道樹單株清查與表型，但須就台灣常見樹種、戶外大尺度與森林林分重新驗證，且現階段限於單株。
Risk of overclaiming勿將本文描述為完整森林或農業數位孿生系統；它是單株植株神經三維重建的前端方法，明言僅適用單株、延伸到田間尺度性能受限，且驗證限於室內柑橘苗。

與 Jacky 博論 / Review 的用途

博士論文支撐博論在數位孿生前端三維重建層的技術選型，說明 NeRF 類神經重建如何透過 ROI 取樣、位姿自動校正與雜湊編碼，在無需去背與受控光照下達成公釐級單株重建，是後續接上感測與碳估算前的資料來源環節。
TJFS Review在 TJFS review 的 Ch7 數位孿生章節，作為以 NeRF 做單株高保真重建並明確指向農業數位孿生資料層的應用實證，與 Tian 等 2025 的林分尺度 NVS 比較、Buonocore 等的概念框架對照，補強由概念到單株、再到林分的技術連續性。
可引用句候選2025 年，Wu 等人發表的文獻中指出，所提出的 Object-Based NeRF 以相機成像軌跡作為先驗自動校正位姿、並以聚焦目標的光線取樣策略免去人工去背，可在 30 秒內完成複雜植株的高保真三維重建，樹高、葉長與葉寬的決定係數分別達 0.9933、0.9881 與 0.9883，為農業數位孿生與高通量表型提供前端資料基礎。
不可用來主張勿用本文單獨作為「森林數位孿生已落地」的證據，也勿宣稱該方法已驗證於森林林分尺度、戶外大場景或台灣樹種；其結論明言僅適用單株、延伸到田間尺度性能受限。

授權與圖表重用

Article licenseCC-BY-4.0
Figure reuse policyDO_NOT_REUSE_ORIGINAL_FIGURES_PUBLICLY_UNTIL_LICENSE_CHECKED
Notes原文首頁版權聲明確認為 open access、Creative Commons Attribution License（CC BY），Frontiers in Plant Science。CC BY 允許在標註原作者與出處下重製與改作；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