{"id":"843bd38a-b75f-4fa3-9009-4b0ba9a4729f","arxiv_id":"2412.05728","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A YOLO11-CBAM model trained on mixed dormant and canopy season images segments apple tree trunks and branches, but year-round generalization is only qualitatively demonstrated.","lead":"This paper combines the YOLO11 object detector with an attention module called CBAM to segment apple tree trunks and branches in orchard images from different seasons. The authors train on winter and summer images and claim the model can work year-round, but the quantitative evidence for that claim is thin.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The year-round generalization claim rests on validation sets never shown to be disjoint from training; Section II-B's split arithmetic (859/2070/84/85) is internally inconsistent with the 78+78 validation images used in Section II-E, so reported metrics may be inflated.","rationale":"The reader's weakest-assumption analysis correctly identifies the 78-image validation sets as the load-bearing point. My stress-test sharpens this into a concrete, checkable internal inconsistency: the split arithmetic in Section II-B (859 images, 2070 training, 84 validation, 85 test) cannot be reconciled with the 156 seasonal validation images used in Section II-E, and the paper never asserts disjointness between training and validation. Because the four-season results are qualitative only, the reported two-season validation metrics are the sole quantitative evidence for the year-round claim. Even setting aside the split problem, the canopy-season numbers are poor (branch mask precision 0.516, trunk 0.643, with low mAP), which by themselves undercut 'effective year-round segmentation.' The authors did collect and manually annotate a substantial field dataset and provide a clear application motivation, which deserves credit, but no code or data are released and the central quantitative claim is not independently auditable. I therefore agree with the reader's REJECT verdict and would not change it; the concern strengthens the case for rejection rather than introducing a new direction.","tokens_in":21770,"tokens_out":3437,"duration_ms":37445,"concrete_test":"Require the authors to release a complete data manifest with image identifiers and split assignments, then compute exact-match and perceptual hashes for all training, validation, and test images. Re-run YOLO11m-seg with and without CBAM, and the best reported configuration, on only those validation images that have no hash match to any training image and that were not captured in the same continuous image sequence as a training image. If the disjoint-subset precision drops materially relative to Tables II and III, or if the 0.83 vs 0.80 CBAM advantage disappears or reverses, the central generalization and CBAM-improvement claims are not supported by the current evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, that YOLO11-CBAM trained on mixed dormant/canopy images can segment trunks and branches year-round, depends entirely on the validity of the validation metrics in Tables II and III. That validity is not established. Section II-B states that 859 images were amassed (553 canopy, remainder dormant), that augmentation 'outputted thrice,' and that the dataset was expanded to 2070 training images with 84 validation and 85 test images. These numbers do not reconcile: 859 × 3 = 2577, not 2070, and 84 + 85 = 169, not the 156 images (78 + 78) described in Section II-E as the dormant and canopy validation sets. The paper never states whether the 78-image validation sets are disjoint from the 859-image training set, nor whether the 84/85 split in Section II-B is the same split used for the seasonal validation. The phrase in Section II that validation images 'had been previously labeled' is consistent with the validation images being drawn from the same annotated pool as training, but no explicit exclusion is stated. If any of the 78+78 validation images overlap with training images—or were captured from the same trees in the same pass—then the precision, recall, and mAP values in Tables II and III, including the headline 0.83 vs 0.80 trunk-precision comparison for YOLO11m-seg, are not unbiased estimates of generalization. Because the four additional seasons (pre-bloom, bloom, thinning, harvest) are evaluated only qualitatively with no reported metrics, the two-season validation is the only quantitative support for the year-round claim. This makes the dataset-split ambiguity load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript reports an instance segmentation system for apple tree trunks and branches, built by inserting Convolutional Block Attention Modules after each convolutional layer of five YOLO11 configurations. The authors train on a mixed dataset of dormant and canopy season images (reported as 859 images before augmentation), validate on 78 images from each of those two seasons, and show example segmentation outputs for four additional seasons. They report precision, recall, and mAP@50 for all configurations and claim the YOLO11-CBAM approach 'demonstrated the potential... to effectively detect and segment tree trunks and branches year-round across all seasonal variations.'","tokens_in":22166,"tokens_out":5823,"duration_ms":57883,"significance":"If the central claim were supported, the work would be a useful application result for orchard robotics, since a single model that works across seasons would reduce the need for season-specific training. The paper's strengths are the real-world data collection in a commercial orchard, the systematic comparison across five YOLO11 sizes, and the inclusion of inference-speed measurements. However, the manuscript does not provide code, data, error bars, or statistical tests, and the quantitative evidence is limited to two seasons; the four additional seasons are only discussed qualitatively. The record is not sufficient to establish the claimed year-round generalization.","major_comments":[{"comment":"The dataset description is internally inconsistent. Section II-B states that 859 images (553 canopy and the remainder dormant) were augmented and that 'each training example [was] outputted thrice,' yielding 2070 training images; 859 × 3 = 2577, not 2070. The same section reports 84 validation and 85 test images (169 total), while Section II-E states that validation used 78 dormant and 78 canopy images (156 total). The manuscript never states whether the 78+78 validation images are disjoint from the 859-image training set. The sentence in Section II that validation images 'had been previously labeled' is consistent with overlap with the annotated training pool. If any validation images come from the same trees, the same imaging passes, or the same annotated pool as training, then Tables II and III are not unbiased estimates of generalization, and the central claim of the paper is not supported. The split arithmetic and the disjointness condition must be resolved before the reported metrics can be trusted.","section":"Section II-B and II-E"},{"comment":"The canopy-season validation, one of only two quantitative seasonal evaluations, shows weak performance. For YOLO11s-seg, branch mask mAP@50 is 0.34 and trunk mask mAP@50 is 0.525; for YOLO11m-seg, all-class mask mAP@50 is 0.319. These numbers are substantially below the dormant-season values in Table II (for example, YOLO11m-seg all-class mask mAP@50 is 0.886 there). The abstract's phrase 'effectively detect and segment ... year-round' is not quantitatively supported by these canopy results, which are also reported without confidence intervals or error bars. A claim of year-round efficacy needs either a much stronger canopy result or an explicit discussion of why these lower values still count as effective for the intended robotic tasks.","section":"Section III-C, Table III"},{"comment":"The four additional seasons (pre-bloom, flower bloom, green fruit thinning, harvest) are evaluated only descriptively. No precision, recall, or mAP values are reported for these seasons anywhere in Section IV; the evidence consists of example images in Figures 16 and 17. Qualitative examples cannot establish the 'across all seasonal variations' portion of the abstract's claim. The authors should either supply quantitative metrics for these four seasons or substantially scale back the generalization claim to two-season validation plus qualitative exploration.","section":"Section II-F and Figures 16-17"},{"comment":"The headline CBAM benefit, 0.83 versus 0.80 trunk mask precision for YOLO11m-seg, comes from the training-validation phase (Table I), not from the held-out seasonal validation sets in Tables II and III. No standard deviation, confidence interval, or significance test is reported for this comparison or for any other CBAM versus non-CBAM comparison. The differences are as small as 0.02-0.03, and Figure 14 only shows YOLO11m-seg; there is no evidence that CBAM consistently improves the other configurations. Without repeated runs or cross-validation, the paper does not establish that CBAM provides a reliable improvement.","section":"Section III-A and Figure 14"},{"comment":"The conclusion concedes that the dataset is 'relatively modest' and recommends substantial expansion plus image registration between seasons as future work. This concession is in tension with the abstract's assertion of demonstrated year-round potential. At best, the study is a proof-of-concept on two seasons with qualitative examples from four others. The central claim must be reworded to match the evidence, or the missing quantitative multi-season results must be supplied.","section":"Section V, Conclusion"}],"minor_comments":[{"comment":"Equation (3) is labeled MIoU and is not used in the results; mAP@50 is never formally defined, and Equation (4) writes '109' where it should be 10^9, given the surrounding text about giga floating-point operations.","section":"Section II-D, Eqs. (1)-(4)"},{"comment":"In the YOLO11x-seg row for the Branch class, box precision is reported as 0.1066 and mask precision as 0.199; these values are far out of line with all other rows and appear to be typographical errors that must be corrected.","section":"Table III"},{"comment":"The paragraph beginning 'This model was further subjected to testing across four distinct seasonal variations...' appears twice verbatim, once before Section IV-A and once near the end of the Discussion; one copy should be removed.","section":"Section IV"},{"comment":"The abstract states that 'the highest recall and precision metrics were observed in the YOLO11x-seg-CBAM and YOLO11m-seg-CBAM respectively' without specifying whether these are training-phase or validation-phase results; this ambiguity should be resolved in the abstract.","section":"Abstract"},{"comment":"The manuscript does not include a data or code availability statement, which makes it difficult for readers to verify the dataset split and the reported single-run metrics; adding such a statement would improve reproducibility.","section":"General"}],"recommendation":"reject","confidential_remarks":"The central claims of the paper are not supported by the evidence as presented: the dataset split is arithmetically inconsistent, the disjointness of validation from training is not established, the canopy-season metrics are weak, the four additional seasons have no quantitative evaluation, and the CBAM improvement is a single-run comparison with no uncertainty assessment. The paper could potentially be revised into a two-season proof-of-concept if the split issues are fixed and the claims are scaled back, but in its current form the load-bearing issues are too fundamental for the stated conclusions. I would also ask the editor to check the heavy reliance on the authors' own preprints as architectural references for YOLO11; those citations may be legitimate, but they are not an independent specification of the model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a decent, well-motivated engineering study that overclaims the year-round generalization. The real contribution is the mixed dormant+canopy training set for trunk/branch instance segmentation in apple orchards, with a five-model comparison and CBAM ablations. That is incremental but not worthless; collecting and labeling 859 images across two seasons in a commercial orchard is real work. The dormant-season validation numbers are strong (trunk mAP@50 of 0.96 on a modest held-out set). The paper also honestly reports the canopy-season weakness: branch mAP@50 tops out at 0.34.\n\nThe largest soft spot is that the paper never clearly states whether the 78-image seasonal validation sets are disjoint from the training images. The text gives three different split descriptions: 84/85 train/val/test in II-B, 78+78 seasonal validation in II-E, and an 8:1:1 ratio in Figure 6. The stress-test note's arithmetic about 859×3=2577 is actually resolvable (690 training images × 3 = 2070, with 84+85 held out), but the 78+78 vs 84/85 discrepancy remains unexplained. If any of those validation images came from the same trees in the same pass as training, the reported metrics are inflated and the central claim is not supported. Second, the four non-dormant/canopy seasons are shown only as qualitative examples; no numbers. That makes \"year-round\" a hypothesis, not a result. Third, the CBAM gains over plain YOLO11 are 0.01-0.03 in most classes, reported without error bars or significance tests; they may be noise. Fourth, no code or data is released, and several architecture details cite the authors' own unreviewed preprints rather than a stable source.\n\nWho is this for: agricultural robotics readers who want a baseline for apple-tree trunk/branch segmentation and a dataset recipe. It is not a methodological advance. A competent referee could turn this into a solid archival paper by forcing the authors to clarify the split, present per-season metrics for all six seasons, and add variability estimates.\n\nRecommendation: send to peer review, expecting major revision. The data collection is real, the topic is relevant, and the flaws are fixable.","headline":"Useful field data and a reasonable baseline, but the year-round claim outruns the evidence due to dataset-split ambiguity and missing metrics for four of six seasons.","tokens_in":22621,"tokens_out":4197,"would_cite":false,"duration_ms":41503,"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 YOLO11-CBAM model trained on dormant and dense-canopy orchard images can segment tree trunks and branches in every tested season of the apple-growing year.","keywords":["YOLO11 instance segmentation","CBAM attention module","apple orchard","tree trunk segmentation","branch segmentation","multi-season vision","agricultural robotics","deep learning"],"falsifier":"Compute the overlap between the 859-image training pool and the two 78-image validation sets, then test the trained model on images from an orchard row or season that supplied no training images; if validation precision drops sharply under that fully held-out split, the claimed year-round generalization does not hold.","tokens_in":21599,"feed_emoji":"🍎","tokens_out":5409,"duration_ms":51295,"temperature":0.7,"pith_summary":"The paper tests whether a single instance-segmentation model can recognize apple tree trunks and branches year-round, rather than building a separate model for each season. It fuses the Convolutional Block Attention Module (CBAM) into five YOLO11 configurations and trains them on a mixed dataset of fully dormant, leafless trees and dense summer canopy. The model is then validated on separate dormant and canopy image sets and qualitatively tested during pre-bloom, flower bloom, fruit thinning, and harvest. The clearest CBAM benefit appears in training metrics, where YOLO11m-seg reaches 0.83 mask precision on trunks versus 0.80 without CBAM, and 0.75 versus 0.73 on branches. If the validation sets are truly independent, the result suggests orchard robots could run one continuously operating perception model instead of season-specific detectors.","feed_headline":"One model segments apple tree trunks in every season","feed_subtitle":"Mixed dormant-and-canopy training plus an attention module lifts YOLO11 precision on trunks and branches year-round.","key_machinery":"The load-bearing mechanism is the Convolutional Block Attention Module (CBAM), a two-stage attention module that first weights feature channels by global average and max pooling through a multi-layer perceptron, then weights spatial regions by pooling across channels and passing the result through a convolution. CBAM is inserted after each convolutional layer in all five YOLO11 configurations (n, s, m, l, x), with the intended effect of making the network concentrate on the channels and spatial areas that distinguish trunks and branches from training poles, trellis wires, and dense foliage. The mixed-season training set itself—859 images from January (leafless) and June (full canopy), augmented to 2070 training images—is the second piece of machinery, since the paper's hypothesis is that these two extremes bracket all other seasons.","core_discovery":"On the paper's own terms, the central claim is that training on the two extreme seasonal states of an apple orchard—complete dormancy and high-density canopy—equips a YOLO11 instance-segmentation model to detect and segment trunks and branches across all intermediate seasonal conditions. The evidence is configuration-dependent: during training, YOLO11m-seg with CBAM achieves the highest mask precision of 0.83 for the trunk class; in dormant-season validation, YOLO11x-seg reaches 0.91 overall mask precision; and in canopy-season validation, YOLO11s-seg leads with 0.516 branch and 0.64 trunk mask precision. The paper presents these results as demonstrating 'potential' for year-round segmentation, with the CBAM integration consistently nudging precision upward compared to the same YOLO11 models without it.","pith_inferences":["Beyond the paper: the strongest test of the extreme-season training hypothesis would be a two-sided ablation—training on each season alone and showing that performance drops on the other season—which is not reported here.","Beyond the paper: Section II-B reports an 84-image validation and 85-image test split, while Section II-E uses 78-image seasonal validation sets; the paper never states how those sets relate to the 859-image training pool, so a reader should treat the reported generalization numbers as upper bounds until the disjointness of validation from training is demonstrated.","Beyond the paper: the paper's own future direction of registering dormant-season images to canopy-season counterparts for each tree could turn the model from a trunk-and-branch segmenter into an always-available structural map of the orchard, making thinning and harvesting decisions possible even when foliage hides branches.","Beyond the paper: a direct field test on an entirely new orchard row or a different commercial orchard, with images that were never part of training or validation, would settle whether the reported precision reflects season generalization rather than orchard-specific memorization."],"forward_implications":["If the result holds, orchard robots can deploy one continuously running trunk-and-branch perception model instead of swapping season-specific detectors.","CBAM's precision gains on trunks (0.83 vs 0.80) and branches (0.75 vs 0.73) in YOLO11m-seg indicate that attention modules give a low-cost accuracy boost without redesigning the detector.","Reliable branch masks give robots the limb cross-sectional area needed for automated green-fruit-thinning decisions and for collision-free pruning and harvesting navigation.","Small configurations such as YOLO11n-seg, with about 3.0 ms inference time, suggest real-time operation is within reach for field robots.","Testing across pre-bloom, bloom, thinning, and harvest seasons, even if qualitative, supports the claim that mixed-season training transfers to conditions unseen during training."],"supporting_citations":[{"why":"Supplies the channel-and-spatial attention module that the paper inserts after every convolutional layer of YOLO11.","marker":"[59]"},{"why":"Defines the YOLO11 instance-segmentation architecture that serves as the base model for the CBAM fusion.","marker":"[56]"},{"why":"Documents YOLO11-based perception in orchard settings, motivating the choice of YOLO11 as the backbone for this task.","marker":"[58]"},{"why":"Earlier CNN-based semantic segmentation of apple tree trunks and branches that this work extends to multi-season instance segmentation.","marker":"[43]"},{"why":"Establishes the difficulty of segmenting obscured tree branches, the central challenge the mixed-season approach addresses.","marker":"[44]"},{"why":"Connects branch diameter to limb cross-sectional area for crop-load estimation, the downstream use that makes branch segmentation valuable.","marker":"[62]"}],"fun_headline_variants":["Attention module sharpens YOLO11 apple tree part masks","Year-round apple tree segmentation with YOLO11-CBAM","YOLO11 plus attention segments orchard trees in any season","Precision gain in apple branch segmentation via CBAM","One model for all seasons: apple trunk and branch segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result stands only if the 78 dormant-season and 78 canopy-season validation images were not also used in training and fairly represent what the model will see in other orchards and other years.","fun_headline_variants_meta":{"raw":{"variants":["Attention module sharpens YOLO11 apple tree part masks","Year-round apple tree segmentation with YOLO11-CBAM","YOLO11 plus attention segments orchard trees in any season","Precision gain in apple branch segmentation via CBAM","One model for all seasons: apple trunk and branch segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000677,"raw_usage":{"total_tokens":3155,"prompt_tokens":1099,"completion_tokens":2056,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":715,"completion_tokens_details":{"reasoning_tokens":1972}},"tokens_in":715,"tokens_out":2056,"duration_ms":15270,"temperature":1.0,"reasoning_tokens":1972,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:24:45.187795+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the overlap between the 859-image training pool and the two 78-image validation sets, then test the trained model on images from an orchard row or season that supplied no training images; if validation precision drops sharply under that fully held-out split, the claimed year-round generalization does not hold.","supporting_citations":[{"cited_title":"YOLO11 and Vision Transformers based 3D Pose Estimation of Immature Green Fruits in Commercial Apple Orchards for Robotic Thinning","cited_arxiv_id":"2410.19846","evidence_quote":"Documents YOLO11-based perception in orchard settings, motivating the choice of YOLO11 as the backbone for this task."},{"cited_title":"Apple tree trunk and branch segmentation for automatic trellis training using convolutional neural network based semantic segmenta- tion,","cited_arxiv_id":null,"evidence_quote":"Earlier CNN-based semantic segmentation of apple tree trunks and branches that this work extends to multi-season instance segmentation."},{"cited_title":"Obscured tree branches segmentation and 3d reconstruction using deep learning and geometrical constraints,","cited_arxiv_id":null,"evidence_quote":"Establishes the difficulty of segmenting obscured tree branches, the central challenge the mixed-season approach addresses."},{"cited_title":"Machine Vision-Based Crop-Load Estimation Using YOLOv8","cited_arxiv_id":"2304.13282","evidence_quote":"Connects branch diameter to limb cross-sectional area for crop-load estimation, the downstream use that makes branch segmentation valuable."}],"review_version":1}