{"id":"fa494a64-dd29-4dfa-98e1-602a5077877c","arxiv_id":"2505.10923","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GrowSplat aligns 3D Gaussian splats from different days to create temporal digital twin videos of plants growing in a greenhouse.","lead":"A team from UC Berkeley and NPEC built a pipeline that combines 3D Gaussian Splatting with point-cloud registration to create time-lapse 3D models of growing plants from multi-view greenhouse cameras. It lets plant scientists watch entire growth cycles in 3D without cutting or touching the plant.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pipeline never estimates the non-rigid deformation field d_tk from Eq. (1), so the claimed temporal consistency rests on an untested assumption that rigid FGR+Colored ICP alignment suffices at 1.4-2.4 day growth intervals.","rationale":"The reader's weakest assumption identifies the same gap: the method solves only the rigid alignment part of Problem 1 while the problem statement requires a non-rigid deformation field. My stress-test confirms this is the most load-bearing issue, because every downstream claim about temporal consistency and phenotyping traits depends on successive point clouds being aligned accurately. The paper provides no quantitative registration error, no baseline comparison to non-rigid methods such as Chebrolu et al., and no ablation isolating the effect of filtering steps; Section VII explicitly defers quantitative evaluation to future work. A conditional verdict is appropriate: the engineering contributions (3DGS with segmentation masks, FPFH/FGR plus Colored ICP) are plausible and the visual results are encouraging, but the central 4D-consistency claim should not be accepted without a test that separates rigid alignment error from unmodeled non-rigid growth. The proposed synthetic deformation experiment would settle whether the rigid-only pipeline is adequate at the growth rates in Table I. If the residual remains small at realistic deformation amplitudes, the concern is resolved; if it does not, the paper should be revised to include a non-rigid refinement step or restrict its claims accordingly.","tokens_in":8338,"tokens_out":4909,"duration_ms":52032,"concrete_test":"Run a controlled synthetic experiment on one reconstructed Sequoia point cloud: apply a known rigid transform plus a known non-rigid deformation field (e.g., an Embedded Deformation graph with amplitude scaled from 0 to 5 cm, mimicking leaf expansion and bending) to create a time t+1 cloud; run GrowSplat's FPFH/FGR + Colored ICP pipeline; measure (i) error in recovering the known rigid component and (ii) residual point-to-plane distance after alignment as a function of deformation amplitude. If residual grows with amplitude and exceeds the typical leaf-thickness scale at the Table I growth rates, the rigid-only method cannot substantiate the claimed 4D consistency.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is a consistent 4D temporal digital twin over 40-55 time steps. The problem statement explicitly models growth as T_tk(P_tk) = T_tk P_tk + d_tk (Eq. 1), with d_tk a non-rigid deformation field, and Problem 1 optimizes over both T_1:K and d_1:K (Eq. 2), including the constraint d_tk^T d_tk <= beta (Eq. 4). The proposed method, however, only estimates the rigid component: downsampling, FPFH/FGR coarse alignment, and Colored ICP fine alignment all recover an SE(3) transform. No step estimates d_tk, and no experiment measures the residual left by unmodeled non-rigid growth. For Sequoia (Avg Delta_t = 2.4 days) and Quinoa (Avg Delta_t = 1.4 days), young plants can visibly change leaf angle, curvature, and area between acquisitions; a rigid alignment will fold that non-rigid change into the recovered pose and leave systematic overlap error, producing apparent drift or ghosting in the temporal twin. Section VII concedes that quantitative evaluations are future work and that biological priors are needed to reduce physically impossible transformations. The qualitative figures (Figs. 4, 5, 8) are suggestive but cannot establish temporal consistency, and no code or data are released for independent checks. Thus the load-bearing assumption is not that FGR/ICP works in general, but that rigid alignment alone is sufficient at these growth rates; this is exactly the assumption the paper never tests.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents GrowSplat, a pipeline for constructing temporal digital twins of plants from multi-view images captured by the NPEC Maxi-Marvin system. At each time step, a 3D Gaussian Splatting model is trained using Splatfacto-MCMC, and the resulting point clouds are aligned across time using FPFH-based Fast Global Registration followed by Colored ICP. The method is demonstrated on two datasets (Sequoia, 40 steps; Quinoa, 55 steps) with qualitative results only. The paper states that quantitative evaluation is future work.","tokens_in":8710,"tokens_out":5385,"duration_ms":49688,"significance":"If the claimed temporal consistency were validated, the system could enable non-destructive, high-throughput plant phenotyping. The integration of 3DGS with existing registration tools is straightforward, and the paper does not introduce new algorithmic components. The significance is currently limited by the absence of any quantitative evaluation, baselines, or ablations, and by the mismatch between the stated non-rigid registration problem and the rigid-only method. The qualitative results are suggestive but not sufficient to establish the central claim.","major_comments":[{"comment":"The manuscript provides no quantitative evaluation: no registration error, no reconstruction quality metric, no comparison against baselines, and no ablation of the filtering or registration steps. Section VII explicitly states that 'We will also carry out quantitative evaluations of GrowSplat', which confirms that the central claim of a 'consistent 4D model' is not yet supported. This is a load-bearing omission that prevents verification of the paper's main contribution.","section":"Section VII, Limitations and Future Work"},{"comment":"The problem statement in Eq. (1) defines the transformation as T_tk(P_tk) = T_tk P_tk + d_tk, and Problem 1 optimizes over both T_1:K and d_1:K subject to the constraint d_tk^T d_tk ≤ β. However, the proposed method in Section V-B recovers only the rigid SE(3) component via FPFH/FGR and Colored ICP; no step estimates the non-rigid deformation field d_tk, and no experiment measures the residual error caused by unmodeled growth. At the given average intervals (2.4 and 1.4 days), leaf expansion and stem bending can be significant. The paper does not test the assumption that rigid alignment alone yields a coherent 4D model, so the central claim is unsupported.","section":"Section IV-C, Eqs. (1)-(4); Section V-B"},{"comment":"The conversion of the division-model radial distortion coefficient κ to Nerfstudio's polynomial model is stated as K1 = -κ/√(w^2+h^2), K2 = K1^2, without derivation or validation. If this approximation is inaccurate, the reconstructed 3D point clouds will contain systematic geometric distortion, which propagates into the registration and final temporal model. The paper should justify this conversion or validate it with reprojection error or equivalent.","section":"Section III-B"},{"comment":"The method relies on several free parameters (filtering thresholds, lighting factor, α and β in constraints (3)-(4)), but their values are never specified, and the constraints (3)-(4) are not enforced in the algorithm. This makes the method difficult to reproduce and means the claimed solution to Problem 1 is not actually implemented. Additionally, no code or data are released for independent verification.","section":"Section V-A, V-B, and IV-C"}],"minor_comments":[{"comment":"The notation T_tk is used both as a function and as a matrix; this is confusing. Consider using distinct symbols, e.g., M_tk for the matrix and f_tk for the function.","section":"Section IV-C, Eq. (1)"},{"comment":"The phrase 'we use segmentation masks that segment the plants from the rest of the image' is redundant; simpler phrasing would improve readability.","section":"Section V-A"},{"comment":"The 'lighting factor' is mentioned but is never defined or quantified; please provide details on its value and how it is applied.","section":"Section V-A"},{"comment":"The word 'consists' is misspelled as 'consits' in the caption.","section":"Figure 2 caption"},{"comment":"It would be helpful to include the number of images per time step and the image resolution, as these directly affect reconstruction quality and the feasibility of the approach.","section":"Table I"}],"recommendation":"reject","confidential_remarks":"This manuscript appears to be a preliminary workshop-style report rather than a complete journal submission. The lack of any quantitative evaluation, the mismatch between the stated non-rigid problem and the rigid method, and the explicit deferral of evaluation to future work make the central claim unsupported. The authors may be encouraged to resubmit after substantially extending the work with quantitative results and an explicit treatment of non-rigid deformation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nGrowSplat is a clean system description: it uses 3D Gaussian Splatting on multi-view greenhouse images, then registers the splat point clouds across time with FPFH/Fast Global Registration and Colored ICP, and renders the aligned models as a 4D growth sequence. The qualitative results for Sequoia (40 timesteps) and Quinoa (55 timesteps) show visible, plausible growth. That integration is real and the codebase choices are sensible. I give credit for the honest limitations section—the authors state clearly that quantitative evaluation is future work.\n\nThe central problem is that the paper's own problem statement (Eqs. 1–4) includes a non-rigid deformation field d_tk, but the proposed method only estimates rigid SE(3) transforms. Leaf expansion, stem bending, and curvature changes between 1.4 and 2.4 day intervals are exactly the non-rigid motions that a rigid registration will partially absorb or leave as misalignment. No metric measures that residual. There is no registration error, no reconstruction error, no baseline comparison (Li et al. 2013 and Chebrolu et al. 2020 both exist), and no ablation of the filtering steps or the lighting factor. So the central claim—a temporally consistent 4D digital twin—is untested.\n\nThese are load-bearing gaps, not minor polish. They are also addressable, and the authors have already identified most of them. The paper is a reasonable workshop demonstration, but I would not send the current version to a journal referee. I would not cite it as evidence of a working method. If I needed a pointer to the application of 3DGS in plant phenotyping, it is a legitimate related-work cite.\n\nTo make it a full paper, I would ask for: a quantitative registration error on held-out timesteps (e.g., RMSE to manually aligned data or a non-rigid baseline), a comparison against existing 4D plant registration, an ablation of the splat filtering and the lighting factor, and an explicit statement of why rigid-only alignment suffices—or an estimate of d_tk. Releasing the Maxi-Marvin data would also go a long way.\n\nI would engage with a revision. As it stands, I see it as an early-stage system paper, suited for a workshop, not a serious referee.","headline":"GrowSplat is an honest but unvalidated system paper: the non-rigid deformation in the problem statement is never estimated, and temporal consistency is shown only qualitatively.","tokens_in":9293,"tokens_out":4081,"would_cite":false,"duration_ms":42285,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GrowSplat claims a fixed multi-camera rig plus coarse-to-fine rigid registration of Gaussian-splat point clouds produces temporally consistent 4D models of plant growth.","keywords":["4D plant modeling","temporal digital twin","3D Gaussian splatting","point cloud registration","plant phenotyping","Fast Global Registration","Colored ICP","non-rigid growth"],"falsifier":"Take two scans of the same plant two to three days apart, run GrowSplat, and measure the residual nearest-neighbour distance between registered point clouds separately for young leaves and for the main stem; if leaf residuals are systematically larger and grow with time, the rigid-registration assumption is failing at the leaf level. A second check is to label a leaf tip in the earlier cloud and see whether its registered position lands within a leaf of the later cloud.","tokens_in":8139,"feed_emoji":"🌱","tokens_out":7584,"duration_ms":74863,"temperature":0.7,"pith_summary":"GrowSplat sets out to turn ordinary multi-view photographs of a plant, taken day after day by a fixed camera rig, into a single coherent 4D model of the plant's development. The core move is to reconstruct each time step as a 3D Gaussian splat, extract its point cloud, and then align successive point clouds in a common frame using a coarse-to-fine registration pipeline: fast point feature histograms with Fast Global Registration, followed by Colored ICP. If the claim holds, plant phenotyping—measuring leaf angle, internode length, growth rate, and other traits—can be done non-destructively and at industrial scale, since the same scanned plant can be followed over its whole growth cycle. The paper demonstrates the pipeline on two species with 40 and 55 time steps respectively, with an average interval between scans of 1 to 2.5 days. It presents visual results and leaves quantitative trait evaluation as future work.","feed_headline":"Gaussian splats build 4D twins of growing plants","feed_subtitle":"A fixed camera rig plus two-stage point-cloud registration tracks Sequoia and Quinoa across 40 to 55 time steps.","key_machinery":"The load-bearing mechanism is the two-stage rigid registration applied to Gaussian-splat point clouds. Fast point feature histograms (FPFH), a local geometric descriptor invariant to rigid motion, provide the coarse feature matches; Fast Global Registration turns those matches into an initial rigid transform; Colored ICP refines the transform using both geometry and color. The 3D Gaussian splat is the representation that makes the point clouds dense and detailed enough for FPFH-based correspondence search to work.","core_discovery":"The central claim is that a complete 4D digital twin of a growing plant can be assembled from a sequence of 3D Gaussian splat reconstructions, one per observation day, aligned into a common plant frame by a coarse-to-fine rigid registration pipeline. For each time step the method trains a Gaussian splat from the 15 multi-view images, filters and downsamples the resulting splats, computes fast point feature histograms, runs Fast Global Registration with RANSAC rejection for coarse alignment, and refines with Colored ICP. The paper argues the result is a temporally consistent sequence of point clouds and rendered views that captures growth of Sequoia (40 time steps) and Quinoa (55 time steps) well enough to support non-destructive, longitudinal phenotyping. This turns the temporal plant modeling problem into a pipeline that runs on industrial-scale data from a fixed multi-camera setup.","pith_inferences":["In my reading, the biggest open question the paper does not settle is whether the rigid alignment actually tracks leaves through time: if a young leaf grows more than the alignment tolerance between two scans, the 'same' leaf may appear as a different structure in the later cloud, and this could be tested with leaf-level correspondences.","One natural extension, implied but not implemented, is to estimate the deformation field from the problem statement using non-rigid registration or a learned growth model; that would turn the visual temporal twin into one with stable leaf identity.","If the registration is as consistent as the visuals suggest, interpolating between registered time steps would yield smoother growth trajectories and could fill in days when the plant was not scanned, something the paper lists as future work.","The paper's own Section VII says biological priors such as monotonic leaf growth and quantitative evaluations are future work; until those exist, 'temporal consistency' remains a visual property rather than a measured one."],"forward_implications":["If GrowSplat's alignment holds, breeders can extract longitudinal trait measurements like leaf angle and internode length directly from registered 4D models, replacing manual and destructive sampling.","Because the pipeline uses fast Gaussian-splat training and a fixed multi-camera rig, it can be rolled out to scan many plants in sequence, making high-throughput phenotyping practical.","The discrete time-step 4D model can serve as the backbone for downstream tasks such as biomass estimation, leaf detection, and growth-rate analysis.","Adding a new observation day reduces to training one splat and running registration, so the 4D model can be updated incrementally over a full growth cycle.","The method's dependence on rigid registration implies it will work best when scans are close enough in time (average under about three days) that growth between steps is small."],"supporting_citations":[{"why":"Supplies the 3D Gaussian splatting representation at the core of each time-step reconstruction.","marker":"[21]"},{"why":"Provides the FPFH local descriptors used to find putative correspondences during coarse alignment.","marker":"[31]"},{"why":"Supplies Fast Global Registration, the robust optimization that turns FPFH matches into an initial rigid transform.","marker":"[32]"},{"why":"Gives the reconstruction framework (with its default splat trainer) used to turn the multi-view images into Gaussian splats.","marker":"[16]"},{"why":"Provides the MCMC-based strategy that densifies and positions the splats during training.","marker":"[30]"},{"why":"Describes the fixed 15-camera capture system and calibration used to collect the image sequences.","marker":"[26]"}],"fun_headline_variants":["GrowSplat builds 4D plant twins from multi-view images","Gaussian splatting gives plants a 4D time-lapse twin","Plant growth in 4D: Gaussian splat temporal twins","Sequoia and Quinoa tracked in 4D with Gaussian splats","4D digital twins of growing plants via Gaussian splats"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes rigid alignment alone can put successive scans of a growing plant into a common frame, even though leaves expand and stems bend between scans, and the paper does not measure how large those non-rigid changes are relative to the alignment error.","fun_headline_variants_meta":{"raw":{"variants":["GrowSplat builds 4D plant twins from multi-view images","Gaussian splatting gives plants a 4D time-lapse twin","Plant growth in 4D: Gaussian splat temporal twins","Sequoia and Quinoa tracked in 4D with Gaussian splats","4D digital twins of growing plants via Gaussian splats"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1448,"prompt_tokens":878,"completion_tokens":570,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":477}},"tokens_in":494,"tokens_out":570,"duration_ms":5308,"temperature":1.0,"reasoning_tokens":477,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:00:11.481734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two scans of the same plant two to three days apart, run GrowSplat, and measure the residual nearest-neighbour distance between registered point clouds separately for young leaves and for the main stem; if leaf residuals are systematically larger and grow with time, the rigid-registration assumption is failing at the leaf level. A second check is to label a leaf tip in the earlier cloud and see whether its registered position lands within a leaf of the later cloud.","supporting_citations":[{"cited_title":"Fast global registration,","cited_arxiv_id":null,"evidence_quote":"Supplies Fast Global Registration, the robust optimization that turns FPFH matches into an initial rigid transform."},{"cited_title":"3d gaussian splatting as markov chain monte carlo,","cited_arxiv_id":null,"evidence_quote":"Provides the MCMC-based strategy that densifies and positions the splats during training."},{"cited_title":"3D imaging – the Maxi-MARVIN","cited_arxiv_id":null,"evidence_quote":"Describes the fixed 15-camera capture system and calibration used to collect the image sequences."}],"review_version":1}