{"id":"28f24395-689e-4643-a5fe-83ab00483f84","arxiv_id":"2507.23487","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An RGB-D pipeline using YOLOv8-Seg, CycleGAN inpainting, tilt correction, and polynomial regression estimates strawberry mass with reported errors of 8.11% for isolated and 10.47% for occluded berries.","lead":"This paper describes a camera and software system that estimates the weight of strawberries in the field even when leaves or other fruit obscure part of the berry. The system uses three standard tools, a segmentation network, an image-completion network, and a formula that turns area into volume, and reports average weight errors of about 8 to 10 percent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 8.11%/10.47% mass-error claims rest on an in-sample fit of the area-to-volume polynomial (Eq. 6); no held-out evaluation is reported, so the headline accuracy is not yet supported.","rationale":"The reader correctly identifies the unvalidated depth-fill heuristic in Section II-C as a weak point for occluded cases. I agree that is a real concern. However, I find an even more fundamental issue that threatens both the occluded and isolated headline numbers: the area-to-volume polynomial (Eq. 6) is fit and evaluated on the same field samples. The paper describes a 6:3:1 split only for the 1200 images used in training YOLOv8-Seg and CycleGAN; the regression step is described as fitting a cubic to 'measured samples' and then immediately reporting prediction errors in Tables II–III. No test split or cross-validation is mentioned for the regression. With four parameters and R^2=0.904, an in-sample evaluation can substantially understate out-of-sample error, especially if the sample is small or the area range is limited. Since the central claim is specifically the quantitative accuracy values, this concern is load-bearing. The concrete test I propose, a held-out evaluation or leave-one-out CV of Eq. 6, would settle it. If the errors remain around 8–10% on held-out data, the paper's central claim is supported; if they inflate, the current experiments overstate accuracy. I therefore keep the reader's CONDITIONAL verdict: the concern is fixable by re-analysis and does not invalidate the pipeline's design, but it must be addressed before the accuracy figures can be accepted.","tokens_in":9598,"tokens_out":5235,"duration_ms":52434,"concrete_test":"Split the measured field samples into fitting and evaluation subsets (or use leave-one-out cross-validation) for the cubic polynomial in Eq. 6, then recompute the mean absolute percentage errors for isolated and occluded strawberries using only held-out samples. If the isolated error rises from 8.11% to, say, >15% while the variance also increases, the central claim of the paper is not supported by the reported experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section III-C, the authors state: 'To assess the accuracy of the algorithm... Field samples were collected... we fitted a cubic polynomial regression model' (Eq. 6) and then report average errors for isolated and occluded strawberries in Tables II and III. The 6:3:1 split described in Section II-A applies to the 1200 RGB-D images used for YOLOv8-Seg and CycleGAN training, not to the polynomial regression, which is fit on the measured field samples. There is no indication that the samples in Tables II–III are disjoint from those used to fit Eq. 6 or that any cross-validation was performed. Evaluating a 4-parameter cubic on its training data is expected to underestimate true prediction error, so the headline '8.11% for isolated, 10.47% for occluded' may not reflect performance on new strawberries. This is the most load-bearing concern because both headline numbers, and the comparison with the literature, depend on these errors being unbiased. The depth-fill heuristic (Section II-C) is also unvalidated, but even the isolated-strawberry number, which does not use it, is subject to the in-sample problem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a pipeline for non-destructive, online mass estimation of table-top grown strawberries in field conditions, targeting both isolated and occluded fruits. The pipeline consists of YOLOv8-Seg instance segmentation, CycleGAN-based occlusion completion, a geometric tilt-angle correction of the projected frontal area, cubic polynomial regression from cross-sectional area to volume, and conversion to mass using an average strawberry density. Experiments reported mean mass estimation errors of 8.11% for isolated and 10.47% for occluded strawberries, with CycleGAN outperforming LaMa on pixel area ratio and IoU metrics for occlusion recovery. The paper is evaluated on real field data and includes visual comparisons, but the central accuracy claims rest on an area-volume regression that appears to be fitted and evaluated on the same samples.","tokens_in":10006,"tokens_out":3409,"duration_ms":35348,"significance":"If the reported accuracies are reproducible on independent data, the proposed system would be a valuable practical contribution to automated harvesting and yield monitoring in strawberries, particularly because it addresses occlusions with CycleGAN and accounts for pose variations. The use of genuine field-captured RGB-D data and the direct comparison with a state-of-the-art inpainting method (LaMa) are strengths, and the paper includes useful visual verification. However, the current evidence does not yet support the headline error rates: the polynomial regression in Eq. (6) is fit on field samples, and the errors in Tables II–III appear to be computed on the same samples, making them in-sample residuals rather than out-of-sample predictions. The depth-fill heuristic and constant-density assumption are also not quantitatively validated. No code or data are provided, further limiting independent verification.","major_comments":[{"comment":"The cubic polynomial regression in Eq. (6) is fit to field samples, and the average errors in Tables II and III are computed on the same experimental data. The 6:3:1 split described in Section II-A applies to the 1,200 images used for YOLOv8-Seg and CycleGAN training, not to the regression samples. There is no statement that the field samples were split into fitting and evaluation subsets, nor any cross-validation. Consequently, the reported 8.11% and 10.47% mean mass errors are likely the residuals of the fit rather than predictions on new strawberries, and the headline accuracy claim in the abstract is not yet supported. Please report the number of field samples, perform a train/test split or cross-validation for the polynomial, and report the out-of-sample errors.","section":"III-C, Eq. (6), Tables II and III"},{"comment":"In occluded scenarios, the missing depth behind the restored region is filled using 'the average depth of isolated strawberries.' This is a constant-fill heuristic, and no validation or comparison against a no-fill baseline, a per-pixel depth model, or an alternative interpolation is provided. Since this fill is applied to all occluded cases, the reported 10.47% occluded mass error is confounded with the error of this assumption. An ablation that removes or varies the depth-fill is necessary to establish that the occlusion-completion pipeline itself brings the claimed benefit.","section":"II-C"},{"comment":"The mass is calculated by multiplying the estimated volume by 'the average density of ripe strawberries,' but the paper never reports the actual density value, how it was measured, or its variability. Section IV acknowledges that density changes with variety, maturity, and moisture content, so a constant density is a potential source of systematic error. Please report the measured density (with units), the number of samples used to obtain it, and a sensitivity analysis showing how the reported mass errors would change within the observed density range.","section":"III-C"}],"minor_comments":[{"comment":"The text states that the measured data points and fitted curves are shown in Fig. 10, but the regression plot is Fig. 9; Figure 10 shows field test images. Please correct the cross-reference.","section":"III-C"},{"comment":"The last paragraph of the Discussion contains the typo 'masss estimation'; it should be 'mass estimation.'","section":"IV"},{"comment":"The abstract and the main text state that in the IoU interval [0.9–1], LaMa achieves 47.7%, while the caption of Fig. 6 says LaMa reaches only 39.7%. These numbers are inconsistent and should be reconciled.","section":"Abstract, III-B, Fig. 6"},{"comment":"The notation in Table I, such as '1 + 0.112' and '1 - 0.13,' is unconventional and ambiguous. Please report the mean and standard deviation (or variance) explicitly, and define how the deviation from the ideal value is computed.","section":"Table I"},{"comment":"The paper reports a validation mAP@0.5 of 0.91 for YOLOv8-Seg but does not state whether the validation set used for this metric is part of the 6:3:1 split described for the 1,200 images, and the numbers of images in the training, testing, and validation subsets are not given.","section":"II-A"},{"comment":"The paper does not state the number of field samples used to fit Eq. (6) and to compute the errors in Tables II and III. Reporting the sample size is necessary to assess the statistical reliability of the error rates.","section":"III-C"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the apparent in-sample evaluation of the area-volume regression. If the authors can provide a proper out-of-sample evaluation (e.g., a clean train/test split or cross-validation for the polynomial), the paper could become acceptable. The depth-fill and density assumptions are secondary but also need quantitative support. I would require these revisions before considering the paper for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the pipeline is sensible and the components are well chosen, but the two headline numbers rest on a polynomial fit that appears to be evaluated on the same field samples it was fitted to. No held-out split is described for the regression in Section III-C, so the 8.11% and 10.47% mean errors are likely training residuals, not out-of-sample predictions.\n\nWhat is actually new: applying CycleGAN to complete occluded strawberry regions and then using the completed mask with a tilt-angle correction for volume/mass. The CycleGAN vs LaMa comparison on 325 test images with PAR and IoU is a real evaluation, even if those metrics are not linked to the final mass error. The segmentation result (mAP 0.91) is solid. Credit where due: the Discussion admits the polynomial regression is sensitive to data distribution and that a fixed density ignores variety, maturity, and moisture. That is the right kind of honest limitation statement.\n\nThe soft spots, in proportion. The load-bearing problem is the in-sample fit. Field samples are collected, Eq. 6 is fitted, and Tables II–III report the resulting errors. Without a separate test split or cross-validation, the mass errors are not trustworthy. This affects both the isolated and occluded numbers because both go through the same cubic. The depth-fill heuristic (average depth of isolated strawberries) is crude and unvalidated; the paper does not bound the error it introduces. There is also no no-restoration baseline, so we cannot see what CycleGAN actually buys for mass estimation. And the IoU/PAR comparison, while a genuine ablation of sorts, is disconnected from the mass metric; a PAR of 0.978 and IoU of 92.3% are good, but we don't know how those translate into the 10.47% error.\n\nMinor: \"online\" is not demonstrated end-to-end; the per-frame times (110–200 ms) are plausible but there is no full-pipeline latency. The 256x256 resolution cap on CycleGAN is acknowledged as a detail-fidelity limit.\n\nVerdict: a useful engineering paper for people building strawberry harvesters or yield monitors. It deserves peer review, but only after the authors do a proper held-out or cross-validated evaluation of Eq. 6, add a no-restoration baseline, and report error bars or at least sample counts. If those numbers hold on independent data, this is a solid applied contribution. I would send it to review with the expectation of major revision rather than desk-reject.","headline":"A sensible engineering pipeline for strawberry mass estimation, but the headline 8.11%/10.47% errors may be in-sample fit residuals, not out-of-sample predictions.","tokens_in":10376,"tokens_out":2755,"would_cite":false,"duration_ms":26456,"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":"A camera-only pipeline estimates strawberry mass with mean errors of 8.11% for isolated fruit and 10.47% for occluded fruit, using segmentation, CycleGAN completion, and tilt correction.","keywords":["strawberry mass estimation","RGB-D camera","CycleGAN","occlusion completion","YOLOv8-Seg instance segmentation","tilt-angle correction","polynomial regression","non-destructive fruit weighing"],"falsifier":"Scan occluded strawberries with a method that captures their true hidden geometry, such as multi-view reconstruction or weighing immediately after picking, and compare against the volume produced by the CycleGAN-plus-average-depth pipeline; if mass errors correlate with occlusion fraction or with the depth difference between occluded and isolated berries, the fill heuristic is the culprit.","tokens_in":9403,"feed_emoji":"🍓","tokens_out":6723,"duration_ms":60875,"temperature":0.7,"pith_summary":"This paper tries to establish that strawberry mass can be estimated online, without touching the fruit, from a single RGB-D camera view even when berries are partly hidden by leaves, stems, or neighboring fruit. The proposed pipeline segments each berry, regenerates the hidden portion with a CycleGAN, corrects the frontal projection for the berry's tilt, and maps projected area to volume through a cubic polynomial. Against field-collected samples the mean mass error is 8.11% for isolated strawberries and 10.47% for occluded ones. If those figures hold, mass-based grading and robotic harvesting could be driven by vision alone.","feed_headline":"Camera alone weighs field strawberries to 10 percent","feed_subtitle":"Vision pipeline completes occluded berries and estimates mass to ~10%, enabling grading without touching fruit.","key_machinery":"Three components carry the argument: (1) CycleGAN, an unpaired image-to-image translation network whose cycle-consistency loss lets it map occluded strawberry images to complete ones without paired training data; (2) a pose estimator that locates the stem, tip, and belly convexity of the berry, fits a plane to local convex points, derives the tilt angle $\\theta$ of the central axis, and multiplies the visible area by $\\cos\\theta$ to recover the frontal projection; and (3) a cubic polynomial regression from maximum cross-sectional area to volume, fitted on field samples with $R^2 = 0.9037$, combined with the average density of ripe strawberries to yield mass. The polynomial absorbs the nonlinear area-to-volume relation that makes smaller berries particularly sensitive to area error.","core_discovery":"The central claim is that occlusion need not prevent accurate mass estimation: YOLOv8-Seg isolates each fruit, a CycleGAN completes the occluded region with cycle-consistency training between occluded and intact image domains, and a symmetry-based pose correction converts the visible projection into a true frontal area. A cubic polynomial $y = -24.9926 + 7.1919x - 0.3063x^2 + 0.0052x^3$ then predicts volume from that area, with mass following from average strawberry density. The paper reports mean mass errors of 8.11% for isolated and 10.47% for occluded strawberries, and an occlusion-recovery stage that beats the LaMa inpainting model in pixel-area ratio (0.978 vs 1.112) and in the fraction of reconstructions with IoU in [0.9, 1] (92.3% vs 47.7%).","pith_inferences":["The average-depth fill for occluded regions is the most fragile step: the paper never verifies that an occluded berry's hidden depth equals the mean depth of isolated berries, so systematic distance differences would bias volume and hence mass.","Because mass is volume times a fixed average density, variety, ripeness, and moisture content will shift the error; a per-class or per-ripeness density calibration is a natural extension.","The segment-complete-correct-regress skeleton should transfer to other irregular, mutually occluding fruits such as raspberries or cherry tomatoes, where the same depth-fill question would arise."],"forward_implications":["One RGB-D frame suffices for mass-based grading: the paper's grade thresholds (A > 30 g, B 20-30 g, C 10-20 g, D ≤10 g) become computable in real time at 110-200 ms per frame.","Occlusion no longer forces either geometric shape guessing or destructive weighing, because the CycleGAN stage handles hidden regions that kite-geometry and symmetry-plane predecessors cannot.","The tilt correction removes a systematic source of underestimation for naturally slanted berries, which is consistent with the isolated-fruit area error staying at 4.14%.","The error structure across stages (area 4-6%, volume 7-10%, mass 8-10.5%) identifies where each part of the pipeline loses accuracy, giving a clear target for further improvement."],"supporting_citations":[{"why":"Supplies the unpaired image-to-image translation architecture with cycle-consistency loss that the paper adapts to complete occluded strawberry regions.","marker":"[19]"},{"why":"The RGB-D plus machine-learning strawberry mass estimation work that motivates combining depth features with learned regression for non-destructive weighing.","marker":"[9]"},{"why":"A point-cloud completion method for strawberry weight estimation that the paper's occlusion-recovery stage is compared against conceptually.","marker":"[10]"},{"why":"Symmetry-plane-based 3D fruit shape completion, the predecessor whose instability on deformed RGB-D point clouds motivates the image-domain CycleGAN approach.","marker":"[13]"},{"why":"Kite-geometry calyx occlusion recovery, the limited earlier attempt at reconstructing partly hidden strawberries that the proposed method generalizes.","marker":"[12]"},{"why":"The autonomous strawberry-harvesting robot whose field occlusions define the real-time, non-destructive mass estimation requirement.","marker":"[3]"}],"fun_headline_variants":["Vision pipeline weighs occluded strawberries to 10% error","Deep learning pipeline estimates occluded strawberry mass to 10%","CycleGAN fills gaps to weigh strawberries within 10%","RGB-D plus CycleGAN yields berry mass error near 10%","Vision system overcomes occlusions to estimate strawberry mass"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire error budget rests on the assumption that the depth hidden behind an occluding leaf, stem, or berry can be replaced by the average depth of isolated strawberries, an approximation the paper states but never validates or bounds.","fun_headline_variants_meta":{"raw":{"variants":["Vision pipeline weighs occluded strawberries to 10% error","Deep learning pipeline estimates occluded strawberry mass to 10%","CycleGAN fills gaps to weigh strawberries within 10%","RGB-D plus CycleGAN yields berry mass error near 10%","Vision system overcomes occlusions to estimate strawberry mass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000688,"raw_usage":{"total_tokens":3113,"prompt_tokens":932,"completion_tokens":2181,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2098}},"tokens_in":548,"tokens_out":2181,"duration_ms":17789,"temperature":1.0,"reasoning_tokens":2098,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:41:08.593611+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Scan occluded strawberries with a method that captures their true hidden geometry, such as multi-view reconstruction or weighing immediately after picking, and compare against the volume produced by the CycleGAN-plus-average-depth pipeline; if mass errors correlate with occlusion fraction or with the depth difference between occluded and isolated berries, the fill heuristic is the culprit.","supporting_citations":[{"cited_title":"Un- paired image-to-image translation using cycle-consistent adversarial networks","cited_arxiv_id":null,"evidence_quote":"Supplies the unpaired image-to-image translation architecture with cycle-consistency loss that the paper adapts to complete occluded strawberry regions."},{"cited_title":"Strawberry picking point localization ripeness and weight estimation","cited_arxiv_id":null,"evidence_quote":"The RGB-D plus machine-learning strawberry mass estimation work that motivates combining depth features with learned regression for non-destructive weighing."},{"cited_title":"Strawberry weight estimation based on plane-constrained binary division point cloud completion","cited_arxiv_id":null,"evidence_quote":"A point-cloud completion method for strawberry weight estimation that the paper's occlusion-recovery stage is compared against conceptually."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Symmetry-plane-based 3D fruit shape completion, the predecessor whose instability on deformed RGB-D point clouds motivates the image-domain CycleGAN approach."},{"cited_title":"A simple and efficient method for automatic strawberry shape and size estimation and classification","cited_arxiv_id":null,"evidence_quote":"Kite-geometry calyx occlusion recovery, the limited earlier attempt at reconstructing partly hidden strawberries that the proposed method generalizes."},{"cited_title":"An au- tonomous strawberry-harvesting robot: Design, development, integra- tion, and field evaluation","cited_arxiv_id":null,"evidence_quote":"The autonomous strawberry-harvesting robot whose field occlusions define the real-time, non-destructive mass estimation requirement."}],"review_version":1}