{"id":"da35257e-1a4d-4ca9-ab7b-648735d26b10","arxiv_id":"2502.02850","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"RS-YOLOX combines ECA, ASFF, Varifocal Loss, and SAHI with YOLOX and reports mAP gains of about 5 points on three remote sensing datasets, but the evaluation protocol may leak augmented images into the test set.","lead":"This paper bolsters the YOLOX object detector with four existing techniques, attention, feature fusion, a loss function, and an inference slicer, and reports higher accuracy on three aerial image datasets. The result is a modest engineering improvement, not a new detection principle.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.2 augments DOTA-v1.5 and RSOD before the train/test split, so test images can be near-duplicates of training images; the reported mAP gains may reflect data leakage rather than real accuracy.","rationale":"The paper's main contribution is an accuracy claim, and its evidence is Table 7. The augmentation-before-split ordering is not cosmetic: it directly breaks the assumption that test images are unseen, which is precisely the assumption needed to interpret mAP as detection accuracy. The reader identified the same concern, and the manuscript contains no statement that augmentation was restricted to the training split. TGRS-HRRSD uses the official split, but DOTA-v1.5 and RSOD do not. Since the model and preprocessing are not released, the issue cannot be resolved post hoc. I do not see a machine-checked proof, shipped code, or reproducible artifacts that would counterbalance this empirical flaw. Thus the reader's REJECT verdict is appropriate; my pass does not change it.","tokens_in":27150,"tokens_out":5074,"duration_ms":50274,"concrete_test":"Obtain or reconstruct the DOTA-v1.5 and RSOD preprocessing from Section 4.2 and compute the overlap between test images and augmented training images (e.g., by source crop ID, perceptual hash, or pixel-similarity). If any test image has a near-duplicate counterpart in the training pool, leakage is confirmed. Alternatively, rerun the full pipeline with the split performed before cropping and augmentation and compare mAP50-95; if the value drops by more than a small margin, the Table 7 result is inflated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that RS-YOLOX has the 'highest accuracy' on three remote sensing datasets, evidenced chiefly by Table 7. For DOTA-v1.5, Section 4.2 states that images were cropped to 640x640, 'then we used data augmentation techniques to increase the number of small categories', yielding 42,831 images, and only then split 6:2:2 into train/validation/test. For RSOD the same order is stated: augmentation to 1,916 images before the 6:2:2 split. This means the test partitions of two of the three datasets are drawn from the same augmented pool as the training partitions; augmented transformations of a training crop can appear in the test set. The reported mAP50-95 values (78.49% on DOTA-v1.5 and 78.56% on RSOD, versus 73.57% and 76.27% for YOLOX) may therefore be inflated by near-duplicate recognition rather than generalization. Without split-before-augmentation or explicit source-level disjointness, the comparison is not trustworthy. TGRS-HRRSD uses the official fixed split, so this particular flaw does not apply there, but the headline claim depends on all three datasets. No code, weights, or split lists are released, so the leakage cannot be ruled out from the manuscript. The internal latency inconsistencies (e.g., YOLOX+ECA latency 13.26 ms in Tables 1/2 versus 14.46 ms in Table 5) and the lack of error bars are additional but secondary concerns.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RS-YOLOX, a remote sensing object detector obtained by adding Efficient Channel Attention (ECA) to the YOLOX backbone, Adaptively Spatial Feature Fusion (ASFF) to the neck, replacing the confidence loss with Varifocal Loss (VFL), and combining the detector with the SAHI inference framework. The method is evaluated on DOTA-v1.5, TGRS-HRRSD, and RSOD, and the central claim is that RS-YOLOX achieves the highest detection accuracy on these datasets, with Table 7 reporting 78.49% mAP50-95 on DOTA-v1.5 versus 73.57% for YOLOX and analogous gains on the other two datasets.","tokens_in":27440,"tokens_out":3930,"duration_ms":37166,"significance":"The paper is a purely empirical engineering contribution that assembles existing modules (ECA, ASFF, VFL, SAHI) into a YOLOX-based detector; it contains no new architecture or theory. The entire significance rests on the claimed accuracy improvement. If the experimental results were reliable, the combination could be useful for practitioners who need a turnkey high-accuracy detector for satellite imagery. However, the evaluation is compromised by a likely train/test leakage in two of the three datasets, by single-run results without error bars, and by an incompletely specified SAHI component. The paper also does not release code or split lists, so the reported numbers cannot be independently verified. These issues are load-bearing because the only claim of the paper is the empirical accuracy claim.","major_comments":[{"comment":"The preprocessing order for DOTA-v1.5 and RSOD introduces a likely train/test leakage. The text states that DOTA-v1.5 images were cropped to 640×640, 'then we used data augmentation techniques' yielding 42,831 images, and only then split into 6:2:2 train/validation/test. For RSOD the same order is described: augmentation to 1,916 images before the 6:2:2 split. Because augmentation is applied to the full set before splitting, augmented versions of a training crop can appear in the test set, so the reported mAP values in Table 7 (78.49% vs 73.57% on DOTA-v1.5, 78.56% vs 76.27% on RSOD) may be inflated by near-duplicate recognition rather than by genuine generalization. The paper does not state that augmented copies were restricted to the training split, and no split lists or code are provided. This directly undermines the central accuracy claim.","section":"Section 4.2"},{"comment":"All accuracy results are reported from a single training run, with no error bars, multiple seeds, or statistical significance tests. The claimed improvements are small in several cases (e.g., the +1.33 mAP for VFL in Table 5, and the +0.07 mAP difference between YOLOX+NAM and YOLOX+ECA in Table 1), and run-to-run variance in training stochastic detectors is typically on the order of a few tenths of a percent. The authors should report the mean and standard deviation over at least three independent runs, or otherwise justify that the differences are not noise.","section":"Section 4.4 and Table 7"},{"comment":"The latency numbers are internally inconsistent. YOLOX+ECA is reported at 13.26 ms in Tables 1 and 2, but the same configuration shows 14.46 ms in Table 5. Similarly, RS-YOLOX (the +VFL row in Table 5) has 17.67 ms latency on DOTA-v1.5, while Table 7 reports 17.86 ms for RS-YOLOX on the same dataset. Because latency is one of the two evaluation metrics claimed in Section 4.3, the paper must specify the exact measurement protocol (warm-up, batch size, precision, input resolution) and reconcile these contradictions, or the speed claims cannot be assessed.","section":"Tables 1, 2, 5, and 7"},{"comment":"The contribution of the SAHI component is not quantified. SAHI is presented in Section 3.5 as a central part of RS-YOLOX, yet Table 7 reports mAP and latency on cropped 640×640 images, where SAHI does not apply, and the table does not state whether SAHI was used for the reported numbers. The slice size, overlap ratio, and NMS settings for the SAHI inference pipeline are not given. The paper should include an explicit ablation with and without SAHI on full-resolution images, and report SAHI hyperparameters, so that the reader can separate the effect of the network modifications from the effect of the inference framework.","section":"Section 3.5 and Table 7"}],"minor_comments":[{"comment":"The phrase 'highest accuracy in detecting objects in remote sensing image datasets' is overbroad; the claim should be qualified to the three evaluated datasets and to the specific evaluation protocol.","section":"Abstract and Section 1"},{"comment":"The formula for the ECA kernel size k contains OCR-like formatting artifacts ('|୪୭୥ మ େ ஓ ൅ ୠ ஓ |୭ୢୢ') and is not readable; the typeset equation should be fixed.","section":"Section 3.2, Equation (2)"},{"comment":"The definition of AP in Equation (7) lacks the integration limits and the variable name in the integrand; the equation should be written as AP = ∫₀¹ P(r) dr.","section":"Section 4.3, Equations (7)-(8)"},{"comment":"There are numerous typographical and grammatical errors (e.g., 'the modelʹs', 'mode' instead of 'model', 'respectivel' in the introduction, and inconsistent unit spacing). A thorough language editing pass is needed.","section":"General"},{"comment":"Reference [19] is a GitHub repository without version or commit information, and reference [58] appears to be incompletely cited; both should be completed or updated.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is an empirical combination paper with no code release. The main risk is that the headline results are compromised by the augmentation-before-split protocol in Section 4.2. If the authors can re-run the experiments with a correct split and release the split lists or code, the paper may become publishable as a useful engineering contribution. The lack of error bars and the SAHI quantification gaps are secondary but must also be addressed. For a journal of this scope, the methodological flaw, not the novelty, is the decisive issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Takeaway: the paper is a cleanly written incremental combination of four known modules (ECA, ASFF, VFL, SAHI) on YOLOX. The authors are honest about provenance, and the ablation design is systematic. But the central accuracy claim is not credible as reported because two of the three datasets appear to leak training augmentation into the test set.\n\nWhat's genuine: the attention comparison (SE, CBAM, NAM, ULSAM, ECA) on DOTA-v1.5 is a useful micro-study. The choice of ECA is justified by the numbers, and the authors correctly note ASFF was already used with YOLOv3 (ref 54), so they don't oversell novelty. TGRS-HRRSD uses the official split, so that result is less exposed.\n\nThe soft spots are serious. Section 4.2 explicitly describes augmenting the full DOTA-v1.5 and RSOD datasets before the 6:2:2 split, giving 42,831 and 1,916 images respectively. There is no statement that augmented copies were kept out of the test partition, and no code or split lists are released. That means near-duplicate recognition can account for part of the reported gains (e.g., 78.49 vs 73.57 on DOTA-v1.5). The 'highest accuracy' claim in the abstract is also overbroad: Table 7 only includes YOLO-family models and one Faster R-CNN, and the paper does not compare against modern oriented detectors or current DOTA leaderboard methods. Secondary issues: latency is inconsistent across tables (YOLOX+ECA 13.26 ms in Table 1 vs 14.46 ms in Table 5), there are no error bars, and SAHI is only shown in qualitative figures without a quantified mAP contribution.\n\nWho this is for: a practitioner might find the module recipe useful, but without code or trustworthy numbers there is little to take away. A serious referee could force a redo of the experiments with split-before-augmentation, multiple seeds, consistent latency, and code release. I would not cite the current numbers. If I were the editor, I would send it to a reviewer rather than desk reject, because the ablation is structured and the leakage is fixable in a revision.","headline":"Incremental YOLOX recipe with a systematic ablation, but the headline accuracy gains are compromised by likely train/test leakage in two of three datasets.","tokens_in":27988,"tokens_out":3424,"would_cite":false,"duration_ms":32962,"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 modified YOLOX detector, RS-YOLOX, outperforms all compared models on three satellite remote sensing benchmarks, making small-object detection in large images more practical.","keywords":["object detection","remote sensing image","YOLOX","Efficient Channel Attention","Adaptively Spatial Feature Fusion","Varifocal Loss","Slicing Aided Hyper Inference","small object detection"],"falsifier":"Apply the same data augmentation only to the training split of DOTA-v1.5 and RSOD, keep the test set untouched, and rerun the evaluation; if mAP50-95 falls materially below the reported 78.49% and 78.56%, the central accuracy claim is inflated by augmented copies leaking into the test set.","tokens_in":26932,"feed_emoji":"🛰️","tokens_out":9006,"duration_ms":75426,"temperature":0.7,"pith_summary":"The paper tries to establish that a modified YOLOX detector, called RS-YOLOX, can detect objects in satellite and aerial remote sensing images more accurately than previous YOLO-family models. The motivation is that remote sensing images are large, contain many small objects, and have imbalanced foreground and background samples, all of which make standard detectors miss targets. RS-YOLOX addresses these problems by adding an efficient channel attention module to the backbone, attaching adaptive spatial feature fusion to the neck, replacing the binary cross-entropy confidence loss with Varifocal Loss, and using a slicing-aided inference step for high-resolution inputs. On DOTA-v1.5, TGRS-HRRSD, and RSOD, the paper reports the highest accuracy among the models it compares, with the main DOTA-v1.5 ablation giving 78.49% mAP50-95 versus 73.57% for the original YOLOX.","feed_headline":"Improved YOLO detector hits 78.5% mAP on satellite images","feed_subtitle":"With four targeted tweaks, the RS-YOLOX detector outperforms every compared YOLO model on three aerial benchmarks.","key_machinery":"Four components carry the argument. Efficient Channel Attention (ECA) replaces the channel-reduction design of squeeze-and-excitation blocks with a one-dimensional convolution over channels after global average pooling, so the backbone emphasizes important channels without adding parameters. Adaptively Spatial Feature Fusion (ASFF) learns per-pixel weight maps for the three PAFPN feature scales, letting small-object features from coarser scales be folded into the fine-scale map. Varifocal Loss (VFL) is an asymmetric loss that weights positive samples by their IoU with the ground-truth box and applies focal-style down-weighting to negatives, countering the background-heavy imbalance of remote sensing scenes. Slicing Aided Hyper Inference (SAHI) cuts a large image into overlapping tiles, runs the detector on each tile, maps the boxes back to the original coordinates, and removes duplicates with non-maximum suppression.","core_discovery":"The central claim is that RS-YOLOX is a high-precision detector for remote sensing images, outperforming all compared one-stage and two-stage detectors on the three chosen datasets. On DOTA-v1.5 the model reaches 78.49% mAP50-95 and 92.83% mAP50, compared with 73.57% and 90.12% for YOLOX; on TGRS-HRRSD it reaches 68.26% mAP50-95 and 88.04% mAP50; on RSOD it reaches 78.56% mAP50-95 and 93.07% mAP50. Here mAP50-95 is the mean average precision averaged over Intersection-over-Union thresholds from 0.50 to 0.95, and mAP50 uses the 0.50 threshold. The ablation shows that each modification contributes: ECA adds 2.01 percentage points, ASFF adds 1.58, and VFL adds 1.33 on DOTA-v1.5. The paper also claims that combining the improved detector with the slicing-aided inference pipeline enables direct, reliable detection on large high-resolution images that miss many objects when fed whole into the model.","pith_inferences":["A testable extension would be applying the same ECA-plus-ASFF-plus-VFL recipe to other one-stage detectors and checking whether the per-component gains are as stable as they are here, since the paper only demonstrates the combination on YOLOX.","The reported superiority is tied to the paper's own cropping and augmentation protocol; comparing against official DOTA-v1.5 leaderboard entries would be a more stringent test of the 'highest accuracy' claim.","Because SAHI works with any detector, the accuracy gains and the large-image handling are separable contributions: a user could adopt only the slicing pipeline with another detector, or only the training-side modifications on already-tiled inputs."],"forward_implications":["If the reported numbers are taken at face value, RS-YOLOX would be the most accurate detector among the compared YOLO-family models on all three datasets, with DOTA-v1.5 mAP50-95 rising from 73.57% to 78.49%.","The ablation indicates that the three training-side changes are additive, so the ECA-plus-ASFF-plus-VFL recipe could transfer to other anchor-free detectors facing small-object and sample-imbalance problems.","SAHI integration means the model can be applied directly to full-size remote sensing images, avoiding the missed detections and false positives shown when a large image is resized and fed whole.","The accuracy gains come with a real cost: parameters on DOTA-v1.5 grow from 8.94M to 14.38M and latency from 13.66ms to 17.86ms, so the method favors accuracy over real-time speed."],"supporting_citations":[{"why":"Defines the YOLOX detector that RS-YOLOX modifies and supplies the main accuracy baseline to beat.","marker":"[20]"},{"why":"Supplies the Efficient Channel Attention module used in the backbone, the source of the first reported accuracy gain.","marker":"[37]"},{"why":"Supplies the Adaptively Spatial Feature Fusion method attached to the neck, the source of the second reported accuracy gain.","marker":"[38]"},{"why":"Introduces Varifocal Loss, which replaces binary cross-entropy as the confidence loss and yields the third reported accuracy gain.","marker":"[39]"},{"why":"Introduces the slicing-aided hyper-inference pipeline used to detect objects in large high-resolution images.","marker":"[40]"},{"why":"Provide the DOTA-v1.5 dataset, the primary benchmark and the source of the main ablation numbers.","marker":"[42-44]"},{"why":"Provide the TGRS-HRRSD dataset used as the second benchmark, including official comparison results.","marker":"[45, 46]"},{"why":"Provide the RSOD dataset used as the third benchmark.","marker":"[47, 48]"},{"why":"Supplies the YOLOv4 comparison baseline for the claim of surpassing previous YOLO models in accuracy.","marker":"[18]"},{"why":"Supply the YOLOv5 comparison baseline, the closest previous YOLO-family competitor in several accuracy columns.","marker":"[19, 24]"}],"fun_headline_variants":["RS-YOLOX: satellite detector beats YOLO by 5 points mAP","Satellite object detection: RS-YOLOX tops three benchmarks","Four tweaks boost YOLO for satellite images to 78.5% mAP","RS-YOLOX: ECA, ASFF, VFL slice their way to top mAP","Improved YOLOX model wins on DOTA, HRRSD, and RSOD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported test accuracy reflects generalization to genuinely unseen images: the paper augments the full DOTA-v1.5 and RSOD datasets and only then splits them into training, validation, and test sets, so if augmented copies of training images leak into the test set, the mAP numbers are inflated by near-duplicates.","fun_headline_variants_meta":{"raw":{"variants":["RS-YOLOX: satellite detector beats YOLO by 5 points mAP","Satellite object detection: RS-YOLOX tops three benchmarks","Four tweaks boost YOLO for satellite images to 78.5% mAP","RS-YOLOX: ECA, ASFF, VFL slice their way to top mAP","Improved YOLOX model wins on DOTA, HRRSD, and RSOD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000571,"raw_usage":{"total_tokens":2725,"prompt_tokens":992,"completion_tokens":1733,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":1623}},"tokens_in":608,"tokens_out":1733,"duration_ms":12064,"temperature":1.0,"reasoning_tokens":1623,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:51:52.734906+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the same data augmentation only to the training split of DOTA-v1.5 and RSOD, keep the test set untouched, and rerun the evaluation; if mAP50-95 falls materially below the reported 78.49% and 78.56%, the central accuracy claim is inflated by augmented copies leaking into the test set.","supporting_citations":[{"cited_title":"ECA‐Net: Efficient Channel Attention for Deep Convolutional Neural Net‐ works","cited_arxiv_id":null,"evidence_quote":"Supplies the Efficient Channel Attention module used in the backbone, the source of the first reported accuracy gain."},{"cited_title":"VarifocalNet: An IoU‐aware Dense Object Detector","cited_arxiv_id":null,"evidence_quote":"Introduces Varifocal Loss, which replaces binary cross-entropy as the confidence loss and yields the third reported accuracy gain."}],"review_version":1}