{"id":"ae0ac3a4-ddba-4abd-8cc0-29227b07a6b6","arxiv_id":"2505.00737","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A review of 3D reconstruction techniques for plant phenotyping, comparing classical methods, NeRF, and 3D Gaussian Splatting on methodology, applications, and future directions.","lead":"This paper surveys how 3D reconstruction methods, from laser scanning to NeRF and 3D Gaussian Splatting, are used to measure plants. It summarizes recent studies and argues these newer methods could make plant measurement faster and more automated.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Incomplete PRISMA reporting, including an 'xx' placeholder for the number of classical papers included, makes the survey's central claim of systematic comprehensiveness unverifiable.","rationale":"The reader's conditional verdict already flags the unfinished appendix and PRISMA reporting, and my stress-test agrees that this is the most serious issue. I did not find evidence that the technical summaries of NeRF or 3DGS are internally wrong: the equations for volume rendering and Gaussian splatting are standard, the described pipelines match the cited works, and the limitations section acknowledges computational costs and dataset scarcity. The paper also has independent support in the form of a public GitHub repository with code and datasets, which gives partial verification of the survey's scope. However, the central claim of systematic comprehensiveness is not verifiable because the methodology section contains an explicit placeholder and lacks the standard PRISMA flow documentation. The reader identified 'selected papers are representative' as a weakest assumption; my concern is more specific: the paper's own Appendix reveals that the selection process is unfinished, so representativeness cannot currently be assessed. This does not move the verdict from CONDITIONAL because the requirement is a concrete fix (completing the PRISMA record and counts) rather than a fundamental flaw in the review's conclusions. If the placeholder were a harmless typo this would be minor, but it directly undermines the one place where the survey substantiates its 'systematic and comprehensive' claim, so it is load-bearing.","tokens_in":22917,"tokens_out":3187,"duration_ms":34289,"concrete_test":"Complete the PRISMA flow diagram and replace the 'xx' placeholder with the actual number of included classical-method papers. Then re-run the stated database queries (Web of Science, ScienceDirect, Springer, IEEE Xplore, arXiv) for 2018.01–2025.02 using the keywords listed in the Appendix, and compare the resulting eligible studies against Table 4. In particular, check whether any 2024–2025 3DGS plant phenotyping paper that cites Kerbl et al. (2023) is missing. If the Table 4 entries match the reproducible search and the classical-paper count is comparable in scope to the NeRF and 3DGS coverage, the comprehensiveness claim is supported; if the search cannot be reproduced or a substantial number of eligible studies are absent, the survey should be reframed as a selective overview rather than a comprehensive systematic review.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that it provides a systematic and comprehensive review of 3D reconstruction techniques for plant phenotyping, with conclusions about the relative promise of NeRF and 3DGS. The only evidence of systematic coverage is the appendix's description of a PRISMA-based search. That description is incomplete: it says 'only xx highly representative papers on traditional methods are included', where 'xx' is a placeholder, and it gives no PRISMA flow diagram or phase-wise counts. This is an explicit, self-admitted gap in the methodology. Because the exact number of classical papers is unspecified, a reader cannot tell whether the selection in Table 4 is representative of the field or an ad hoc subset that may bias the survey's comparative statements (e.g., 'classical methods face challenges' versus '3DGS offers efficiency and scalability'). The sentence '4 studies on NeRF (9 papers) and 3GS (3 Papers) in plant phenotyping were selected' is also internally confusing: it does not state a total number of included studies or clarify how '4 studies' relates to '9 papers'. Without a complete and reproducible record of search results, inclusion decisions, and exclusion counts, the central claim of being systematic and comprehensive cannot be checked. This is a load-bearing concern because it affects the validity of the survey's overall conclusion, not just an editorial detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a survey of 3D reconstruction techniques applied to plant phenotyping, with the stated aim of covering classical active and passive methods, Neural Radiance Fields (NeRF), and 3D Gaussian Splatting (3DGS). It provides mathematical preliminaries for NeRF and 3DGS, a taxonomy of evaluation metrics (pixel-level, geometry-level, trait-specific), and a table of recent representative studies. The paper claims to be a systematic review following PRISMA guidelines and includes a GitHub repository link for additional resources. The central assertion is that NeRF and 3DGS, especially 3DGS, offer efficiency and scalability advantages for automated high-throughput phenotyping, while classical methods face issues of noise, data density, and scalability.","tokens_in":23167,"tokens_out":4176,"duration_ms":41196,"significance":"If the methodological gaps are fixed, this survey would be a timely and useful reference for agricultural and computer-vision researchers entering the area. The descriptions of NeRF and 3DGS pipelines (Equations (1)-(11)) are standard and essentially correct, the metric taxonomy in Table 3 is well organized, and Table 4 compiles recent applications across crops. The paper also gives explicit credit to the emergence of datasets such as PlantGaussian and Splanting. Its comparative conclusions are qualitative rather than quantitative, but that is acceptable for a survey; the main risk to the paper's value is the incompleteness of the reported search methodology, which currently prevents verification of the 'systematic' claim.","major_comments":[{"comment":"The sentence 'only xx highly representative papers on traditional methods are included' contains an unresolved 'xx' placeholder. Combined with the absence of a PRISMA flow diagram and phase-wise inclusion/exclusion counts, this undermines the paper's central claim of a systematic and comprehensive review. A reader cannot determine how many classical papers were considered or whether the selection in Table 4 is representative. Please replace the placeholder with the actual number and add a complete PRISMA-style record: search dates, database-specific query strings, numbers of records identified, screened, excluded, and included.","section":"Appendix, page 14"},{"comment":"The statement '4 studies on NeRF (9 papers) and 3GS (3 Papers) in plant phenotyping were selected' is internally confusing. It does not state whether '4 studies' is the total number of included studies or only the NeRF studies, and it does not explain how 4 relates to the 9 NeRF papers and 3 3DGS papers. Please provide clear, non-ambiguous counts of included studies per method category and a grand total.","section":"Appendix, page 14"},{"comment":"The text states that classical methods are 'as summarized in Table 2' and refers readers to Table 2 for a more comprehensive review, but Table 2 is a summary of existing literature reviews, not a summary of classical reconstruction methods. This cross-reference mismatch weakens the survey's coverage of classical methods and would mislead readers. Either rename and restructure Table 2 to actually summarize classical method categories (with representative references) or correct the textual cross-reference.","section":"Section 3.1, page 7"}],"minor_comments":[{"comment":"The entry 'PatchMatch Multi-View Stere' is missing a final 'o'; it should read 'PatchMatch Multi-View Stereo'.","section":"Nomenclature, page 2"},{"comment":"The abbreviation '3GS' is used inconsistently; the paper elsewhere uses '3DGS'. Please standardize to '3DGS' throughout, including the Appendix.","section":"Appendix, page 14"},{"comment":"The text refers to the framework 'Splants' in Ojo et al. (2024), but the reference list correctly gives the title 'Splanting: 3D Plant Capture with Gaussian Splatting'. Please clarify which term denotes the framework and which denotes the dataset to avoid confusion.","section":"Section 3.3, pages 11-12"},{"comment":"The integration variable in Equation (3) is mismatched: the integral is written with respect to 'du' but the integrand is a function of 's'. It should be 'ds' for consistency with the notation used elsewhere in the paper.","section":"Section 2.2, Equation (3)"},{"comment":"The prose describing RMSE says it quantifies 'the average squared difference between predicted and actual values', but the formula correctly gives the square root of that average. Please align the wording with the formula.","section":"Section 2.4.3, RMSE definition"}],"recommendation":"major_revision","confidential_remarks":"The paper's topic is within the journal's scope and the technical content is serviceable, but the unresolved 'xx' placeholder and the unclear study counts in the PRISMA appendix are substantive reporting gaps that must be fixed before the 'systematic review' claim can be accepted. I would not reject on content grounds; the issues appear fixable within a normal revision cycle."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about arXiv:2505.00737: it's a genuinely useful survey of 3D reconstruction in plant phenotyping, and its systematic-review appendix is unfinished in a way that matters for its central claim.\n\nWhat's good: the paper fills a real gap. Previous reviews cover classical imaging or LiDAR, but not the recent NeRF/3DGS wave. The math for NeRF and 3DGS is standard and correctly presented; equations (1)-(11) are accurate, and the descriptions of training and rendering pipelines are clean. The application table (Table 4) and the discussion of downstream tasks give a newcomer a solid map of the field. The writing is readable and the figure choice is appropriate. I also appreciate the github repository for following up on cited data.\n\nWhere it's soft: the appendix describing the PRISMA-based selection is incomplete. It literally says 'only xx highly representative papers on traditional methods are included'—that's an unresolved placeholder. It also says '4 studies on NeRF (9 papers) and 3GS (3 Papers)' without explaining the relationship between studies and papers, and there's no PRISMA flow diagram. Because the paper's abstract promises a 'comprehensive review' and the introduction calls it 'systematic,' this is not purely cosmetic. A reader cannot verify whether the classical-method selection is representative or cherry-picked, and that selection feeds the comparative statements in Section 3 and 4. The fix is straightforward: fill in the count, add a flow diagram or at least phase-wise numbers, and clarify the study/paper distinction. Minor typos like 'Splants' vs 'Splanting' and '3GS' should be cleaned up.\n\nIs the central argument broken? No. The claim that NeRF and 3DGS are promising for high-throughput phenotyping is supported by the cited applications regardless of the exact PRISMA count. The limitations section is honest about computational cost, dataset scarcity, and evaluation fragmentation. The reported numbers from cited works (R^2, MAPE, etc.) look plausible, though I haven't verified them against the originals.\n\nVerdict: this is a competent survey that deserves peer review, not a desk reject. A serious referee would ask for a completed PRISMA appendix and a few clarifications, then it would be a useful contribution for researchers entering the field. I'd recommend accepting conditional on those fixes.\n\nBest.","headline":"Useful survey of 3D reconstruction for plant phenotyping with correct technical summaries, but the incomplete PRISMA appendix undercuts the claim of systematic coverage.","tokens_in":23650,"tokens_out":2268,"would_cite":false,"duration_ms":22970,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"NeRF and 3DGS push plant phenotyping past LiDAR","keywords":["Plant phenotyping","3D reconstruction","Point cloud","Neural radiance fields","3D Gaussian splatting","Deep learning","Precision agriculture","High-throughput phenotyping"],"falsifier":"Run the surveyed NeRF and 3DGS pipelines on a single standardized benchmark of, say, 100 field-grown maize plants with co-registered LiDAR ground truth; if 3DGS's trait MAPE exceeds 20% where the papers report 10\\textendash 11%, or if NeRF's training time stays over an hour per plant at field scale, the survey's central claim that these methods make high-throughput phenotyping practical loses its empirical support.","tokens_in":22763,"feed_emoji":"🌱","tokens_out":8521,"duration_ms":76991,"temperature":0.7,"pith_summary":"This survey sets out to organize the fast-moving field of 3D reconstruction for plant phenotyping around three families: classical depth-sensing and photogrammetric methods, neural radiance fields (NeRF), and 3D Gaussian splatting (3DGS). Its central claim is that the field is shifting from expensive, noisy, hard-to-scale classical pipelines toward deep-learning reconstructions that work from ordinary multi-view images, with 3DGS emerging as the most promising route because it combines explicit geometry with real-time rendering. The paper assembles representative studies across crops such as apple, rice, maize, cotton, and tomato, reports trait-estimation accuracies in terms of $R^2$, RMSE, MAPE, F1, and IoU, and argues that these methods can make automated high-throughput phenotyping practical. A sympathetic reader should come away with a clear map of which method to choose for which phenotyping task and where the remaining bottlenecks lie.","feed_headline":"NeRF and 3DGS push plant phenotyping past LiDAR","feed_subtitle":"A survey finds: neural radiance fields match scanning accuracy; 3D Gaussian splatting adds real-time speed.","key_machinery":"The organizing axes of the review are two scene representations. NeRF models a scene as a continuous volumetric radiance field, an MLP $F_\\theta(\\mathbf{x},\\mathbf{d})$ that maps each 3D location and viewing direction to a color and density $(\\mathbf{c},\\sigma)$, rendered by numerical integration of transmittance along rays. 3DGS replaces this implicit field with an explicit collection of learnable 3D Gaussians, each carrying a center $\\boldsymbol{\\mu}$, a covariance $\\boldsymbol{\\Sigma} = \\boldsymbol{R}\\boldsymbol{S}\\boldsymbol{S}^T\\boldsymbol{R}^T$, an opacity $\\alpha$, and a color $\\mathbf{c}$, rendered by projecting the ellipsoids to 2D and $\\alpha$-blending them in depth order. This representational contrast drives the survey's central comparison: ray-marched implicit fields give photorealistic geometry at high training cost, while splatted explicit Gaussians give real-time rendering and faster optimization, which is what high-throughput phenotyping demands. The review also uses a tripartite metric scheme\\textemdash pixel-level (PSNR, SSIM, LPIPS), geometry-level (IoU, Chamfer distance, boundary overlap, precision/recall/F1), and trait-level ($R^2$, RMSE, MAPE)\\textemdash to compare studies that otherwise report incommensurable numbers.","core_discovery":"The discovery the paper is trying to establish is that neural 3D reconstruction has reached the point where it can do what classical phenotyping hardware does, with cheaper and more flexible inputs. NeRF, trained only on multi-view RGB images and camera poses, reconstructs plant geometry that matches terrestrial laser scanning to sub-millimeter distances in greenhouse trials (0.865 mm mean error) and yields trait estimates with $R^2$ values from 0.89 to 0.98 for height, leaf area, stem thickness, and fruit volume. 3DGS goes further on the operational axis: because it represents geometry as explicit 3D Gaussians that are projected and $\\alpha$-blended rather than ray-marched, it trains and renders far faster, and early plant studies report cotton boll count and stem length MAPEs around 10\\textendash 11\\% from smartphone-captured images. The survey's comparative conclusion is that 3DGS, not NeRF, is the paradigm most likely to scale to high-throughput field phenotyping, provided remaining issues of occlusion, dataset scarcity, and non-standard evaluation metrics are addressed.","pith_inferences":["A concrete economic consequence the paper leaves implicit: if 3DGS trait errors stay near the reported 10\\% MAPE, a smartphone-video pipeline could replace LiDAR for row-crop phenotyping, reducing sensor cost by roughly two orders of magnitude and making per-plant trait measurement feasible at breeding-program scale.","Because the surveyed NeRF/3DGS plant studies mostly adapt generic models (Instant-NGP, Nerfacto, vanilla 3DGS) with no plant-specific loss terms, a testable extension is that crop-aware priors\\textemdash for example, enforcing botanical consistency of leaf skeletons or stem connectivity\\textemdash would close much of the remaining accuracy gap more cheaply than adding sensors.","The survey's call for standardized metrics implies a concrete benchmark design: fixed camera trajectories over a multi-species set with co-registered LiDAR ground truth, reported as $R^2$, MAPE, and Chamfer distance simultaneously; without it, the field's reported numbers will remain incommensurable and the comparative ranking unverifiable."],"forward_implications":["NeRF-based phenotyping can replace laser scanners for many indoor and greenhouse traits: a single RGB camera or smartphone, multi-view images, and pose estimation yield trait accuracies in the $R^2 \\approx 0.89$\\textendash $0.98$ range reported in the surveyed studies.","3DGS brings real-time rendering and training fast enough for field-scale work, so the bottleneck shifts from reconstruction speed to data acquisition and segmentation quality rather than geometry computation.","Combining 3DGS with foundation segmentation models such as SAM enables organ-level trait extraction (boll count, stem length, leaf area) directly from Gaussian primitives, as the cotton and PlantGaussian studies demonstrate.","Multi-modal and hyperspectral extensions of NeRF/3DGS, the survey's main future direction, would let phenotyping measure functional traits like chlorophyll and stress alongside structure."],"supporting_citations":[{"why":"Introduces NeRF, the implicit volumetric representation that all surveyed NeRF plant reconstructions build on.","marker":"Mildenhall et al., 2021"},{"why":"Introduces 3D Gaussian splatting, the explicit Gaussian representation the survey argues is the scalable paradigm.","marker":"Kerbl et al., 2023"},{"why":"Provides the greenhouse pepper NeRF-vs-scanner comparison with the 0.865 mm mean distance error.","marker":"Zhao et al., 2024"},{"why":"Evaluates NeRF (Nerfacto) against LiDAR ground truth in field conditions, reporting the F1 = 74.6% field result.","marker":"Arshad et al., 2024"},{"why":"PanicleNeRF: smartphone-based rice panicle reconstruction with SAM+YOLO segmentation, F1 = 86.9%, IoU = 79.8%.","marker":"Yang et al., 2024"},{"why":"3DGS cotton boll and stem trait estimation with MAPE near 11% for boll count and 10.45% for stem length.","marker":"Jiang et al., 2025"},{"why":"Splants framework and dataset for 3DGS plant capture across nine species, supporting the survey's claim of 3DGS generality.","marker":"Ojo et al., 2024"},{"why":"PlantGaussian: 3DGS-based cross-time, cross-scene plant visualization and mesh reconstruction with SAM-based segmentation.","marker":"Shen et al., 2025"},{"why":"3DPhenoMVS, the classical SfM tomato pipeline that serves as the baseline the survey contrasts with NeRF and 3DGS.","marker":"Wang et al., 2022"}],"fun_headline_variants":["3DGS outpaces NeRF for field-ready plant phenotyping","Neural 3D scans match LiDAR, then 3DGS makes it real-time","Survey: 3DGS, not NeRF, will scale field phenotyping","From LiDAR to Gaussians: plant 3D goes neural","NeRF nails precision, 3DGS wins the field race"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's comparative conclusions stand on the accuracy and representativeness of the numbers it transcribes from the cited papers; if those reported results are misquoted, cherry-picked, or measured under incompatible protocols, the ranking of classical, NeRF, and 3DGS methods the survey presents would not hold.","fun_headline_variants_meta":{"raw":{"variants":["3DGS outpaces NeRF for field-ready plant phenotyping","Neural 3D scans match LiDAR, then 3DGS makes it real-time","Survey: 3DGS, not NeRF, will scale field phenotyping","From LiDAR to Gaussians: plant 3D goes neural","NeRF nails precision, 3DGS wins the field race"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3508,"prompt_tokens":1073,"completion_tokens":2435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":2335}},"tokens_in":689,"tokens_out":2435,"duration_ms":18652,"temperature":1.0,"reasoning_tokens":2335,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:07:55.689814+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the surveyed NeRF and 3DGS pipelines on a single standardized benchmark of, say, 100 field-grown maize plants with co-registered LiDAR ground truth; if 3DGS's trait MAPE exceeds 20% where the papers report 10\\textendash 11%, or if NeRF's training time stays over an hour per plant at field scale, the survey's central claim that these methods make high-throughput phenotyping practical loses its empirical support.","supporting_citations":[],"review_version":1}