{"id":"ab94bad5-1bee-4932-ad5a-980d9737210f","arxiv_id":"2411.15808","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"LRSAA is an ensemble of YOLOv11 and MobileNetV3-SSD with Poisson disk sampling and EIoU-based NMS that claims improved object detection on large remote sensing images.","lead":"This paper combines two existing object detectors (YOLOv11 and MobileNetV3-SSD) with Poisson disk sampling, a custom NMS metric, and self-generated training data to detect and label objects in large remote sensing images. It reports higher accuracy than several baselines on urban imagery from Tianjin, Shanghai, and Xiamen, but the method description contains errors and the claimed benefit of synthetic data is not supported by the paper's own table.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The coordinate-remapping matrix printed in Sec. III is degenerate and is never reconciled with the nonzero metrics of Table I; without code-level verification the central 'outperforms all models' claim is not reproducible.","rationale":"The paper is an engineering pipeline: train YOLOv11 and MobileNetV3-SSD on XView crops, crop large urban images via Poisson disk sampling, ensemble the detections, map boxes back to the original image, and optionally retrain with synthetic crops. Its headline claim is that LRSAA outperforms all compared detectors on 640-cut and 320-cut images. That claim stands or falls with the correctness of the evaluation pipeline, and the step connecting patch detections to whole-image boxes is precisely where a wrong equation would invalidate every number in Table I. The Sec. III matrix, as typeset, is internally degenerate: with x_max equal to x_min plus the patch offset, all mapped boxes have zero area and cannot pass the IoU > 0.5 threshold described in Sec. IV-B. Since Table I is full of positive values, the implementation must diverge from the printed method; without code verification the reported comparison is not reproducible. This is the same weakness the reader identified, so I agree with that choice. Other concerns are real but secondary: the EIoU formula is incomplete, the synthetic-data claim in Sec. IV-E appears contradicted by Table II (higher original-data fractions yield better results), the 'balance between accuracy and speed' is never supported by speed measurements, and the Tianjin synthetic data may leak into the Tianjin evaluation. However, the remapping issue is the most load-bearing because it attacks the validity of the central comparison itself. In good faith, the release of code and the plausibility of cropping-based detection are positive aspects; if the concrete repository check shows the implemented mapping is correct, the rejection may soften to conditional pending a rewritten methods section and re-verified results. As written, the reader's REJECT stands unchanged.","tokens_in":7072,"tokens_out":6353,"duration_ms":60572,"concrete_test":"Clone https://github.com/anaerovane/LRSAA and locate the function that maps patch detections back to full-image coordinates. Call it with a known patch box, e.g. [10,20,100,80], and a patch offset of (640,640). If the returned box has xmax equal to the mapped xmin (e.g. xmax=650 instead of 740, ymax=660 instead of 720), the Sec. III transform is literally implemented and Table I's IoU-based metrics cannot be trusted. If the returned box is correct, the printed matrix is a typo, and this particular concern fails to land; the remaining question would then be whether the actual implemented mapping is documented anywhere in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core comparison (Table I in Secs. IV-C/D) requires detection boxes produced on 320/640/1280-pixel crops to be mapped back to the original 6400x6400 image before IoU is computed. That mapping is specified in Sec. III by a matrix equation whose third row is [1 0 0 0] and whose fourth row is [0 1 0 0], so a literal reading sets x_max = x'_min + w*nx and y_max = y'_min + h*ny. Every mapped box then has zero width and zero height, forcing IoU to zero and making the accuracy/F1/mAP values in Table I impossible. The paper never acknowledges this, and the nonzero entries thus imply that the implementation does not follow the printed equation. Because no code listing, ablation, or independent verification of the remapping is supplied, the evaluation behind the claim 'The proposed LRSAA model outperforms all other models' is not reproducible from the text. If the repository implements the printed transform, Table I is meaningless; if it implements a different transform, the paper omits the exact step its central claim depends on. Either way this is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LRSAA, an ensemble of YOLOv11 and MobileNetV3-SSD for object recognition and automatic annotation in large-scale remote sensing images. The pipeline splits large images into crops using Poisson disk sampling, detects objects with the two models, fuses results via EIoU-based NMS, maps detections back to the original image, and then retrains on a mixture of original and synthetic data. Experiments compare LRSAA against RetinaNet, SSD VGG, Faster R-CNN, YOLOv11, and MobileNetV3-SSD on 6400x6400 urban images from Tianjin, Shanghai, and Xiamen, reporting accuracy, F1, and mAP. The paper claims that LRSAA outperforms all competing models, especially with 640-cut and 320-cut images, and that adding synthetic data improves performance.","tokens_in":7266,"tokens_out":4359,"duration_ms":38161,"significance":"If the results held, the paper would offer a practical recipe for large-image remote sensing detection with reduced memory requirements, and the public code repository would aid reproducibility. The use of two complementary detectors, Poisson disk sampling for crop selection, and synthetic-data retraining are all reasonable engineering ideas, and the three-city benchmark with manual annotations is a useful evaluation setup. However, the current manuscript does not sufficiently substantiate these claims: the central coordinate-remapping equation is degenerate, the headline comparison is internally inconsistent, and the synthetic-data experiment has a leakage risk. The strengths are the availability of source code and the attempt to evaluate on real 0.6 m resolution urban imagery, but these do not compensate for the missing technical details and unverified correctness of the mapping step.","major_comments":[{"comment":"The matrix equation for mapping detections from crop coordinates back to the original image is degenerate: the third row is identical to the first and the fourth to the second, so a literal reading sets x_max = x'_min + w*nx and y_max = y'_min + h*ny, forcing every mapped box to have zero width and height. If the implementation follows the printed equation, the IoU-based metrics in Table I would be meaningless; if it does not, the paper omits the exact transformation on which the central claim depends. The authors must provide the corrected matrix or the exact code used for remapping.","section":"Section III, coordinate remapping"},{"comment":"The claim that 'The proposed LRSAA model outperforms all other models' is contradicted by Table I: on Xiamen with 640-cut images, YOLOv11 achieves Accuracy 0.794, while LRSAA achieves 0.781; on Shanghai, MobileNetV3-SSD 640-cut achieves mAP 0.62 versus LRSAA 0.70, but on Xiamen LRSAA 320-cut mAP 0.72 is tied with YOLOv11 0.69 and MobileNetV3-SSD 0.63, while accuracy is lower than YOLOv11. The comparison should be stated per metric and per city, and the headline claim must be qualified to the settings where it actually holds.","section":"Section IV-D, Table I"},{"comment":"The conclusion that 'there is a significant trend toward performance improvement as the proportion of synthetic data increases' is not supported by the table: the 80% synthetic row has Accuracy 0.716, lower than the 60% row's 0.718, and mAP 0.66 versus 0.67; the 40% row has F1 0.698, below the 0% row's 0.706. No error bars or significance tests are reported. Furthermore, the synthetic data are generated from automatically annotated Tianjin images and the model is then evaluated on manually annotated Tianjin images, which is a within-city pseudo-label leakage that artificially inflates the apparent benefit. A cross-city evaluation or a clear separation between training and evaluation cities is needed.","section":"Section IV-E, Table II"},{"comment":"The EIoU equation is printed as 'EIoU = 1 - IoU + ρ2 + v' with ρ2 and v never defined, and the relationship between EIoU as a bounding-box regression loss and its use in NMS is not explained. The paper also omits the ensemble fusion weights, the NMS threshold, and the procedure for combining YOLOv11 and MobileNetV3-SSD outputs. These details are essential for reproducing the claimed improvements.","section":"Section III, EIoU and fusion details"},{"comment":"Table I reports accuracy, F1, and mAP without error bars, number of runs, or statistical significance tests, and the training details (epochs, learning rate, batch size, data augmentation, confidence threshold other than 0.25, IoU threshold definition in 'exceeds 50') are absent. The use of classification accuracy with TN for object detection is also not justified, because true negatives are not well-defined for detection tasks. These omissions make it impossible to assess whether the differences in Table I are meaningful.","section":"Section IV-C/D, experimental methodology"}],"minor_comments":[{"comment":"The phrase 'YOLOv11 and MobileNetV3-SSD trainging' contains a typo; 'training' is intended.","section":"Section III, typos"},{"comment":"The header 'Acurracy F1-score mAP' misspells 'Accuracy'.","section":"Table II"},{"comment":"References [15] and [16] both cite the same Cook 1986 paper; one of the two should be removed or replaced with the Bridson reference that the text actually describes.","section":"Section II-B and References"},{"comment":"The text says 'A result is considered accurate if its IoU value exceeds 50'; this should be 'exceeds 0.5' or 'exceeds 50%'.","section":"Section IV-B"},{"comment":"Table III appears to be a set of detection visualizations, but no explanation is given for the color coding or the selection of examples, which limits its usefulness as evidence.","section":"Table III, Appendix"}],"recommendation":"major_revision","confidential_remarks":"The paper is short and preliminary, and several of the issues are presentation-level (typos, missing definitions), but the coordinate-remapping error and the synthetic-data leakage are substantive. I believe the claims are salvageable if the authors re-run or re-present the experiments with a correct mapping, cross-city validation, and error bars, but the current version does not meet the standard for acceptance. The contradictory Table I claim should be caught by the authors before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: don't spend referee time on v4 as written. The core comparison probably rests on a coordinate remapping that the paper prints incorrectly, and the two central claims—'outperforms all' and 'synthetic data helps'—are not supported by the evidence as presented.\n\nWhat's genuinely there: the pipeline is a sensible engineering combination (YOLOv11 + MobileNetV3-SSD ensemble, Poisson disk cropping, EIoU-based NMS, pseudo-label retraining) for high-res remote sensing annotation, a domain where labeling is expensive. Testing on three 0.6m city datasets is real effort, and Table I's message—crop-based detection crushes whole-image detection—is credible and useful for practitioners. The public code URL is a plus, though I didn't run it.\n\nSoft spots, in order of severity. First, the coordinate remapping matrix has its third row equal to its first, so a literal read maps xmax to xmin plus offset and every mapped box has zero area; then Table I couldn't have nonzero IoU. Either the implementation differs from the printed math or Table I is meaningless, and the paper never acknowledges this. Second, the EIoU equation is incomplete: rho^2 and v are named but not defined, and 'EIoU = 1 - IoU + rho^2 + v' is not a loss function as written. Third, Table II contradicts the text: the conclusion says synthetic data monotonically improves everything, but F1 drops at 40% and accuracy doesn't move monotonically. More importantly, the 0%-synthetic row (0.707/0.706) doesn't match the LRSAA 640-cut baseline in Table I (0.791/0.777), so the comparison setup is unclear. Fourth, the synthetic data comes from model-generated labels on Tianjin and the evaluation includes Tianjin; that is a leakage risk the paper doesn't discuss. There are no error bars, no ablations of ensemble/NMS/sampling choices, and no speed measurements, despite the abstract's 'balance between accuracy and speed.'\n\nNet: the idea is plausible and the direction is worth pursuing, but the load-bearing evaluation is not reproducible from the text. I would desk reject the current version and invite a resubmission with corrected equations, the actual code path for coordinate remapping, error bars, and a leakage-aware experimental design.","headline":"Load-bearing coordinate remapping and evaluation contradictions make this engineering pipeline hard to trust as written, though the crop-based detection direction is sound.","tokens_in":7844,"tokens_out":3317,"would_cite":false,"duration_ms":30162,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LRSAA, an ensemble of YOLOv11 and MobileNetV3-SSD with Poisson-disk tiling and EIOU merging, outperforms five standard detectors in automatic annotation of large urban remote sensing images.","keywords":["remote sensing target recognition","automatic annotation","ensemble learning","YOLOv11","MobileNetV3-SSD","Poisson disk sampling","EIOU loss","large-scale imagery"],"falsifier":"Inspect the released code and run it on one 6400-by-6400 annotated urban tile, printing the remapped boxes before any IoU is computed; if a substantial fraction of mapped boxes have zero width or height, the accuracy, F1, and mAP values in the comparison tables cannot be reproduced and the central outperformance claim collapses.","tokens_in":6823,"feed_emoji":"🛰️","tokens_out":9882,"duration_ms":75035,"temperature":0.7,"pith_summary":"The paper tries to establish that a practical pipeline for recognizing and automatically labeling objects in very large remote sensing images can beat standard detectors by combining two existing models rather than designing a new architecture. The LRSAA pipeline cuts large aerial images into tiles with Poisson disk sampling, runs YOLOv11 and MobileNetV3-SSD on each tile, merges the detections using an EIOU-based non-maximum suppression, and remaps the surviving boxes to the original image. On 0.6-meter-resolution urban imagery of three cities, the authors report accuracy and F1 scores near 0.79 with 640-pixel tiles, ahead of RetinaNet, SSD VGG, Faster R-CNN, and the two base detectors; they also report that mixing in a proportion of the model's own auto-annotated results as synthetic training data improves recognition. If correct, the contribution is a reusable recipe for large-scale automatic annotation rather than a new detector.","feed_headline":"Ensemble model LRSAA beats five baselines on urban satellite images","feed_subtitle":"Resampling large aerial photos into smaller tiles and fusing detections lifts accuracy past 0.79 in tests.","key_machinery":"The load-bearing mechanism is a four-step pipeline: (A) Poisson disk sampling partitions a large image into uniformly distributed, minimally overlapping tiles; (B) YOLOv11 and MobileNetV3-SSD are trained on small tiles and detect objects in each tile; (C) the detected boxes are remapped to original-image coordinates by a matrix offset; and (D) the auto-annotated results are mixed back into the training set as synthetic data and the models are retrained. The identity that carries the merging step is EIOU, an intersection-over-union measure augmented with center-point distance and aspect-ratio terms, used here in non-maximum suppression to discard duplicate boxes more selectively. The ensemble step is result-level: the outputs of both detectors are combined before suppression, so the method's gain comes from fusion, tiling, and remapping rather than from a new detector.","core_discovery":"The central claim is that the LRSAA ensemble—YOLOv11 plus MobileNetV3-SSD, with Poisson disk sampling for tile extraction and EIOU-based non-maximum suppression for merging—produces more accurate object detection and automatic annotation on large-scale urban remote sensing imagery than any individual detector or standard single-scale baseline. The paper supports this with comparative tables in which LRSAA at 640-cut achieves the highest average accuracy (0.791) and F1 (0.777) across Tianjin, Shanghai, and Xiamen, and at 320-cut achieves the highest average mAP (0.72). The authors further claim that retraining on a mix of real data and a proportion of automatically annotated synthetic data improves accuracy, F1, and mAP, and that the pipeline reaches this accuracy while keeping computational demand low through tiling.","pith_inferences":["If the coordinate remapping in the released code matches the printed matrix, the IoU-based numbers in the comparison tables would be invalid, because the printed matrix degenerates every mapped box; a code-level check would be the first thing to try.","The 'synthetic data' here is not synthetic imagery but model-generated annotations on real imagery, so the reported improvement is really a self-training or pseudo-labeling effect that could be tested with confidence thresholds.","The design isolates three variables—ensemble, tiling, and EIOU merging—so a simple ablation with a single detector plus tiling and standard NMS instead of EIOU would reveal which component actually carries the gain."],"forward_implications":["If LRSAA works as claimed, automatic annotation of very large remote sensing images can be built from two standard detectors plus tiling, without designing a purpose-built network.","The reported gains at 320- and 640-pixel tiles imply that smaller tiles, not larger context, drive the accuracy improvement, which suits memory-constrained deployment.","Because the ensemble is modular, the authors expect that further detectors can be added without changing the rest of the pipeline, making the framework extensible.","The reported synthetic-data results imply that an auto-annotation system can partially bootstrap its own training data, reducing manual labeling effort."],"supporting_citations":[{"why":"supplies the XView dataset with over one million bounding-box annotations used for initial model training","marker":"[21]"},{"why":"defines YOLOv11, one of the two detectors in the ensemble","marker":"[10]"},{"why":"defines MobileNetV3, the backbone of the second ensemble detector","marker":"[6]"},{"why":"defines the EIOU loss used in non-maximum suppression for merging detections","marker":"[19]"},{"why":"provides the comparison of Poisson disk generation methods underlying the tile-sampling step","marker":"[17]"},{"why":"is the SSD baseline included in the evaluation table","marker":"[4]"},{"why":"is the Faster R-CNN baseline included in the evaluation table","marker":"[22]"},{"why":"is the RetinaNet baseline included in the evaluation table","marker":"[23]"}],"fun_headline_variants":["LRSAA: YOLOv11 and MobileNetV3-SSD fusion lifts remote sensing accuracy","Poisson disk tiling and EIOU merge: LRSAA tops baselines on aerial images","Combining YOLOv11 and MobileNetV3-SSD improves object detection in remote sensing","LRSAA: tiling and fusion yield top detection on urban satellite imagery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire comparison depends on the coordinate remapping from tile to full image being implemented correctly; the matrix printed in the paper would set every mapped box's maximum coordinate equal to its minimum coordinate plus the tile offset, making every mapped box degenerate, and no code-level evidence is provided that the real implementation avoids this.","fun_headline_variants_meta":{"raw":{"variants":["LRSAA: YOLOv11 and MobileNetV3-SSD fusion lifts remote sensing accuracy","Poisson disk tiling and EIOU merge: LRSAA tops baselines on aerial images","Combining YOLOv11 and MobileNetV3-SSD improves object detection in remote sensing","LRSAA: tiling and fusion yield top detection on urban satellite imagery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000512,"raw_usage":{"total_tokens":2425,"prompt_tokens":817,"completion_tokens":1608,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":433,"completion_tokens_details":{"reasoning_tokens":1511}},"tokens_in":433,"tokens_out":1608,"duration_ms":10566,"temperature":1.0,"reasoning_tokens":1511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:52:05.159697+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released code and run it on one 6400-by-6400 annotated urban tile, printing the remapped boxes before any IoU is computed; if a substantial fraction of mapped boxes have zero width or height, the accuracy, F1, and mAP values in the comparison tables cannot be reproduced and the central outperformance claim collapses.","supporting_citations":[{"cited_title":"Howard, M","cited_arxiv_id":null,"evidence_quote":"defines MobileNetV3, the backbone of the second ensemble detector"},{"cited_title":"”Focal and efficient IOU loss for accurate bounding box regression.” Neurocomputing 506 (2022): 146-157","cited_arxiv_id":null,"evidence_quote":"defines the EIOU loss used in non-maximum suppression for merging detections"},{"cited_title":"”A comparison of methods for generating Poisson disk distributions.” Computer Graphics Forum","cited_arxiv_id":null,"evidence_quote":"provides the comparison of Poisson disk generation methods underlying the tile-sampling step"},{"cited_title":"”Attention and feature fusion SSD for remote sensing object detection.” IEEE Transactions on Instrumentation and Measurement 70 (2021): 1-9","cited_arxiv_id":null,"evidence_quote":"is the SSD baseline included in the evaluation table"},{"cited_title":"”Faster R-CNN: Towards real-time object detection with region proposal networks.” IEEE transactions on pattern analysis and machine intelligence 39.6 (2016): 1137-1149","cited_arxiv_id":null,"evidence_quote":"is the Faster R-CNN baseline included in the evaluation table"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the RetinaNet baseline included in the evaluation table"}],"review_version":1}