{"id":"39fa8e2e-d4bc-4146-af90-0f68a0cf2657","arxiv_id":"2508.00900","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A synthetic-data-trained pipeline for locating rose flower centers in 2D and estimating their depth in stereo images is described, with in-simulation F1 up to about 96-100%, but real-world detection is below a YOLOv5 baseline and real depth is untested.","lead":"The paper presents a two-stage 3D perception pipeline for rose-harvesting robots, combining 2D point-based flower detection with monocular, stereo, or template-matching depth estimation trained on synthetic Blender images. It reports strong in-simulation accuracy, but real-world detection underperforms a fine-tuned YOLOv5 baseline and real-world depth accuracy is never measured, so the claimed simulation-to-real bridge remains unproven.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world 3D depth accuracy is never measured; the paper's sim-to-real bridge is asserted for depth, not demonstrated, and the only real 2D signal favors a baseline.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: sim-to-real transfer for depth is assumed, not measured. I agree, and I would keep the REJECT verdict. My stress-test narrows the concern: the 2D transfer is partially tested and shows a gap (real F1 74.4 vs 95.6 synthetic; YOLOv5s beats the proposed detector on real data), but the 3D transfer is completely untested. The paper's own Section 3.2 and Section 5.7 admit that no real depth labels exist. Without such labels, the abstract's 'bridging simulation and real-world' claim cannot be evaluated for the depth component, which is the component that makes the pipeline a 3D perception system rather than a 2D detector. The concrete test I propose is the minimal experiment that would settle the concern: add real depth ground truth to a modest number of stereo pairs and measure end-to-end depth error. If that test were run and passed, the rejection would need to be reconsidered; as written, the claim is under-supported at the point the robot actually depends on. I am not raising a secondary concern about the inference-time table or number inconsistencies because the missing real-depth evaluation is more load-bearing and sufficient to justify the current verdict.","tokens_in":17603,"tokens_out":4041,"duration_ms":48547,"concrete_test":"Acquire real depth ground truth for the existing stereo rig, for example by placing roses or artificial roses at measured distances from 1 to 4 meters (or by aligning an RGB-D/LiDAR point cloud to the stereo images) for 100+ stereo pairs. Run the trained stereo deep module end-to-end: detect near flower centers in the left image, predict their depth, and compute MAE against the measured depths, binned by distance. If real near-flower MAE stays within roughly 2x the synthetic value (about 10 cm at 2 m, i.e., 5% error instead of 3%), the bridge claim survives; if it is substantially larger or detection loses the near flowers on real data, the paper's real-world claim should be downgraded to a synthetic-demonstration-only result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper claims to bridge simulation and the real world for sparse 3D perception of roses. For that claim, the load-bearing premise is that depth estimates trained on Blender synthetic data remain accurate on real rose images. Section 3.2 explicitly states that the depth of flower centers was not recorded in the real dataset, and Section 5.7 says real-world depth metrics are excluded 'due to the absence of acquired depth ground-truth for real-world data.' Consequently, the abstract's 3% depth error is a synthetic-only result (Table 5: NCCoef near-flower L1 = 0.06 m at 2 m), and no quantitative real-world 3D localization result appears anywhere in the paper. The one real-world signal that is measured points in the opposite direction: fine-tuned YOLOv5s achieves F1 = 80.4 on real near flowers, while the proposed stereo and monocular detectors achieve 74.0 and 78.0 (Table 10). This does not refute the method, but it removes independent support for the claim that the 2D stage transfers well. The central claim is not internally inconsistent; it is under-supported at its most load-bearing point: the real-world depth transfer is asserted rather than measured, and the only real-data comparison available favors a baseline detector.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage pipeline for sparse 3D localization of rose centers in agricultural robotics: a point-based 2D detector (U-Net-like architecture with an EfficientNetV2-S encoder) followed by depth estimation, implemented in monocular and stereo variants and compared with a template-matching triangulation baseline. Training relies on a synthetic Blender-generated rose-farm dataset with dense 2D/3D annotations, while a small real-world stereo dataset with 2D-only labels is used for evaluation. The authors report high synthetic F1 scores and depth errors of 3–8% at 2 m range, and they claim that the approach bridges the simulation-to-real domain gap for 3D perception.","tokens_in":17916,"tokens_out":5865,"duration_ms":61897,"significance":"If fully supported, the paper would offer a practical recipe for training agricultural 3D perception systems entirely on synthetic data with minimal real annotation. The Blender-based synthetic dataset with precise 2D/3D labels and the point-based detection formulation are potentially reusable contributions. However, the two load-bearing claims are not established: real-world 3D localization accuracy is never measured, and the paper's own real-world 2D comparison shows a fine-tuned YOLOv5s baseline outperforming the proposed detector. The headline numbers in the abstract are also not consistently recoverable from the reported tables. With these gaps, the contribution reduces to a synthetic-only depth-estimation study whose real-world transfer remains unvalidated.","major_comments":[{"comment":"The abstract's headline F1 of 95.6% (synthetic) cannot be recovered from the per-category results in Tables 2 and 3: monocular near/distant F1 are 96.6/96.3 and stereo near/distant are 95.5/99.8, and no listed averaging procedure produces 95.6. Likewise, the abstract's 74.4% (real) is the stereo near-flower F1 from Table 3, while the monocular near-flower F1 is 78.0 and the fine-tuned YOLOv5s baseline in Table 10 achieves 80.4; the claimed 'superiority of our method' is therefore not supported by the paper's own real-data comparison.","section":"Abstract; Tables 2–3; Table 10"},{"comment":"The paper never evaluates depth estimation on real-world data: Section 3.2 states that 'the depth of flower centers was not recorded' in the real dataset, and Section 5.7 excludes real-world depth metrics 'due to the absence of acquired depth ground-truth for real-world data.' The abstract's '3% at a 2-meter range' is a synthetic-only result (Table 5, NCCoef near flowers). Since the title and abstract claim a bridge from simulation to real-world applications and the paper's central contribution is 3D localization, this missing measurement leaves the key claim unsupported.","section":"Sections 3.2 and 5.7; abstract"},{"comment":"The 'hit rate' of 100% reported for the proposed stereo methods in Table 9 is inconsistent with the recall values in Table 3 (97.6% for near flowers in both stereo rows), and Section 5.2 does not define 'hit rate' as a metric. The comparison table also mixes F1, mAP, and hit rate across studies without a clear common basis; the paper should either reconcile these numbers or remove the comparison.","section":"Section 5.2; Tables 3 and 9"},{"comment":"The abstract credits the 'lightweight deep neural network' with the 3% depth error at 2 m, but Table 5 shows that 0.06 m (3%) is the template-matching baseline (NCCoef), not the deep stereo network, whose near-flower error is 0.096 m (4.8%). The deep method is superior to template matching only for distant flowers (0.13 m vs 0.20 m); the framing of the results should be corrected to attribute the headline figure to the appropriate method.","section":"Sections 5.5–5.6 and Table 5"}],"minor_comments":[{"comment":"The inference times for 3D monocular are reported as 211 ms (CPU) and 893 ms (GPU); since GPU inference should be faster, these values are likely swapped or mislabeled.","section":"Section 5.8, Table 7"},{"comment":"Several typos appear in the related-work section, including 'localizae' and 'by appying'; the text needs copyediting.","section":"Section 2.3"},{"comment":"The caption reads 'Bounding box size (pixles)' — typo — and the heuristic '60/depth' is introduced without motivation or sensitivity analysis.","section":"Equation (12)"},{"comment":"The confidence threshold of 0.51 and the ±5-pixel spatial tolerance for true positives are presented without justification or an ablation study.","section":"Section 5.2"},{"comment":"The Blender citation lacks publication year or last-updated date; please complete the reference.","section":"Reference [43]"}],"recommendation":"reject","confidential_remarks":"The paper's title and abstract overstate the contribution relative to the evidence. The synthetic dataset is a potentially useful asset, but the central sim-to-real claim for 3D localization is untestable from the presented data because real depth ground truth was never collected, and the only real-world 2D comparison favors a baseline. If the authors can collect real depth labels and re-run the evaluation, that would be a different, potentially publishable study; as submitted, the main claim is unsupported and the reported numbers are internally inconsistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this is a competent engineering paper whose central sim-to-real claim is not supported by the reported evidence. The Blender rose-farm dataset is the most valuable part, and the two-stage pipeline is a sensible combination of known components. But the authors never measure real-world depth, and the one real 2D comparison they run favors a fine-tuned YOLOv5s over their point-based detector.\n\nWhat's new: a synthetic rose-farm dataset with 2D/3D annotations, generated in Blender with stereo pairs, heatmaps, and depth maps. That's a real asset for a niche problem, and the synthetic-only depth numbers (3% at 2 m with template matching, 6.5% at 2-6 m with deep stereo) are plausible and reasonably presented. I also credit the authors for comparing against YOLOv5s and including qualitative failure cases.\n\nSoft spots, in order of severity. First, real depth is never measured: Section 3.2 says depth wasn't recorded, Section 5.7 excludes real depth metrics. So the phrase 'bridging simulation and real-world' applies only to 2D detection, and even there the evidence is mixed. Table 10 shows YOLOv5s at 80.4 F1 vs. 74.0 for the stereo point-based detector on real near flowers. That's the only real-signal comparison, and it goes against the paper's superiority claim. Second, the abstract's 95.6% synthetic F1 is not obviously recoverable from the per-category tables; if it's a weighted average the authors should say so, and the 74.4% real figure is the near-flower stereo number, not a headline aggregate. Third, Table 7 lists GPU inference times slower than CPU (893 vs. 211 ms), which is physically implausible unless the measurement includes something unusual. Minor: the YOLOv5 box heuristic (60/depth) is a reasonable compromise but not a full evaluation.\n\nNone of this makes the pipeline useless. The synthetic data recipe and the two-stage detection-plus-depth architecture could help someone building a rose or flower harvesting perception system. The paper is not self-contradictory, and the engineering is honestly described. But the real-world 3D transfer that the title promises is asserted, not measured.\n\nI'd send this to peer review — the dataset and method deserve scrutiny and the flaws are fixable — but the authors need to either collect real depth labels or explicitly narrow the claim to 2D synthetic-to-real transfer. If the numbers stay uncorrected and the GPU table stays as is, it's a reject.","headline":"Solid engineering and a useful synthetic dataset, but the sim-to-real bridge is asserted for depth, not measured, and the only real 2D comparison favors YOLOv5s.","tokens_in":18434,"tokens_out":3221,"would_cite":false,"duration_ms":33279,"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":"The paper argues that rose-harvesting robots can be guided by a synthetic-trained stereo pipeline that detects rose centers and estimates depth, removing the need for real 3D labels.","keywords":["Damask rose","rose harvesting robot","synthetic data","stereo depth estimation","point-based object detection","3D localization","agricultural robotics","domain transfer"],"falsifier":"Collect real rose-field stereo pairs with ground-truth depths (for example, from LiDAR or a rangefinder) across 0.5 to 6 meters, run the trained stereo deep and template-matching pipelines, and compare predicted versus measured depths; the central claim fails if the error at 2 meters is substantially above the reported roughly 6 to 8 centimeters or if real-data detection F1 does not approach the fine-tuned YOLOv5s baseline of 80.4.","tokens_in":17386,"feed_emoji":"🌹","tokens_out":8251,"duration_ms":89118,"temperature":0.7,"pith_summary":"The paper argues that a rose-harvesting robot can locate Damask rose centers in 3D from stereo RGB images without ever needing a labeled real-world depth dataset. The proposed pipeline detects flower centers as points in a three-channel heatmap (near, distant, background) and then estimates depth at those centers using a lightweight neural network, with a template-matching-and-triangulation alternative for comparison. All training uses a 1,000-sample photorealistic Blender dataset with precise 2D and 3D labels; only a small real stereo set with 2D labels is used for evaluation. The authors report 95.6% F1 on synthetic and 74.4% F1 on real images for 2D detection, and 3% depth error at a 2-meter range on synthetic data. The claim matters because synthetic pretraining plus a point-based stereo head could be a cheap, deployable route to precision harvesting, although the real-data detection gap and the absence of real depth labels define the open risk.","feed_headline":"Synthetic-data pipeline: 95.6% F1, 3% depth error for rose centers","feed_subtitle":"Two-stage stereo network detects rose centers and estimates depth, aiming to cut labeling cost for harvest robots.","key_machinery":"The load-bearing object is the three-channel adaptive-Gaussian heatmap, which turns 2D detection into a dense point-prediction problem and encodes depth through kernel sharpness. The stereo depth decoder then reads the detected centers off a dense depth map produced from shared encoder features of the left and right images. Two-stage training (localizing first, freezing the detector, then training depth) and the custom sigmoid parameterization of the depth output are supporting mechanisms that stabilize learning.","core_discovery":"The paper's central claim is that point-based, sparse localization can be coupled with direct stereo depth regression so that both tasks share an encoder and are trained on synthetic data alone. The localization head is a U-Net-style decoder on an EfficientNetV2-S backbone that outputs three heatmap channels; ground-truth heatmaps are Gaussian peaks whose spatial extent shrinks with flower depth, so the network is encouraged to localize close flowers sharply. The stereo depth decoder receives features from both images, which lets correspondence happen implicitly, and is supervised only at flower centers by a smoothed L1 loss. On the synthetic test set, near-flower localization reaches about 95.5% F1 for the stereo model, and stereo deep depth error is 9.6 cm for near flowers versus 6 cm for template matching with triangulation; the authors express these as 5% and 3% relative errors at 2 meters. The paper's own comparison shows the point detector outperforms a fine-tuned YOLOv5s on synthetic data but trails it on real data (74-point versus 80-point F1), which the authors attribute to the domain gap in the synthetic training set.","pith_inferences":["A decisive test the paper leaves undone is measuring real depth: recording LiDAR or manual distances for a subset of the real stereo pairs and comparing the stereo network's predictions would directly test the sim-to-real depth transfer that the title promises.","The reported real-data result suggests the point-based detector's advantage is environment-specific; combining its heatmap head with a stronger feature extractor or adding a small amount of real labeled imagery to training could close the six-point real F1 gap.","The adaptive-Gaussian heatmap that encodes depth-dependent sharpness could be reused for other sparse grasping targets, such as stems, fruit peduncles, or flower buds, not just rose centers.","Because template matching and deep stereo fail in complementary ranges, a confidence-based selector between the two depth estimates is a natural next system-level improvement."],"forward_implications":["If the transfer claim holds, a rose-harvesting robot could be built with a cheap stereo camera and a model trained on Blender images, eliminating LiDAR and real 3D labeling cost.","The near/distant heatmap split means the system is explicitly optimized to grasp nearby blooms, which is the operationally relevant target for a robotic arm.","Because stereo deep regression beats template matching for distant flowers (13 cm versus 20 cm test L1 error) while template matching wins nearby (6 cm versus 9.6 cm), a hybrid matcher could widen the accurate picking range.","The compact models, under five million parameters, are offered as evidence that the pipeline can run on resource-constrained field hardware."],"supporting_citations":[{"why":"This citation supplies the point-based (keypoint) detection paradigm the heatmap head is built around.","marker":"[4]"},{"why":"This citation gives the 'objects as points' formulation that motivates center regression instead of bounding boxes.","marker":"[5]"},{"why":"This citation is the prior Damask rose stereo-harvesting work whose hit rate and depth error the paper compares against.","marker":"[34]"},{"why":"This citation is the fine-tuned YOLOv5s baseline that the proposed detector is compared with on synthetic and real data.","marker":"[39]"},{"why":"This citation is the Blender tool used to create the synthetic rose-farm dataset with 2D and 3D labels.","marker":"[43]"},{"why":"This citation supplies the U-Net encoder-decoder with skip connections used for heatmap prediction.","marker":"[44]"},{"why":"This citation supplies the EfficientNetV2-S backbone pretrained on ImageNet.","marker":"[45]"},{"why":"This citation defines the normalized cross-correlation criteria used in the template-matching depth path.","marker":"[49]"},{"why":"This citation supplies the triangulation formula depth = bf/disparity for stereo depth computation.","marker":"[50]"}],"fun_headline_variants":["Synthetic data trains rose-harvest vision: 95.6% F1, 3% depth error","Two-stage stereo pipeline for rose centers: 95.6% F1, 3% depth","Sim-to-real 3D perception for rose picking: 95.6% F1, 3% error","Sparse localization plus depth regression: 95.6% F1, 3% depth","Rose harvest robots: synthetic training yields 95.6% F1, 3% depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Blender-generated rose-farm images transfer to real fields for both detection and depth; the depth half of that transfer is never measured because the real dataset lacks depth labels.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic data trains rose-harvest vision: 95.6% F1, 3% depth error","Two-stage stereo pipeline for rose centers: 95.6% F1, 3% depth","Sim-to-real 3D perception for rose picking: 95.6% F1, 3% error","Sparse localization plus depth regression: 95.6% F1, 3% depth","Rose harvest robots: synthetic training yields 95.6% F1, 3% depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000342,"raw_usage":{"total_tokens":1926,"prompt_tokens":1030,"completion_tokens":896,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":766}},"tokens_in":646,"tokens_out":896,"duration_ms":9341,"temperature":1.0,"reasoning_tokens":766,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:05:42.875609+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect real rose-field stereo pairs with ground-truth depths (for example, from LiDAR or a rangefinder) across 0.5 to 6 meters, run the trained stereo deep and template-matching pipelines, and compare predicted versus measured depths; the central claim fails if the error at 2 meters is substantially above the reported roughly 6 to 8 centimeters or if real-data detection F1 does not approach the fine-tuned YOLOv5s baseline of 80.4.","supporting_citations":[{"cited_title":"Center- net: Keypoint triplets for object detection,","cited_arxiv_id":null,"evidence_quote":"This citation supplies the point-based (keypoint) detection paradigm the heatmap head is built around."},{"cited_title":"Robotic harvesting of rosa damascena using stereoscopic ma- chine vision,","cited_arxiv_id":null,"evidence_quote":"This citation is the prior Damask rose stereo-harvesting work whose hit rate and depth error the paper compares against."},{"cited_title":"Organization","cited_arxiv_id":null,"evidence_quote":"This citation is the Blender tool used to create the synthetic rose-farm dataset with 2D and 3D labels."},{"cited_title":"U-net: Convolutional net- works for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"This citation supplies the U-Net encoder-decoder with skip connections used for heatmap prediction."},{"cited_title":"Fast r-cnn,","cited_arxiv_id":null,"evidence_quote":"This citation defines the normalized cross-correlation criteria used in the template-matching depth path."},{"cited_title":"Template matching,","cited_arxiv_id":null,"evidence_quote":"This citation supplies the triangulation formula depth = bf/disparity for stereo depth computation."}],"review_version":1}