{"id":"f9d65e31-b75b-4f53-bfd7-bf652b6149af","arxiv_id":"2412.12562","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A lightweight oriented object detector that stacks known modules (wavelet transform, P2-based feature pyramid, dynamic convolution) onto YOLOv8, reporting competitive DOTA/DIOR-R and very high UCAS-AOD mAP, though with internal inconsistencies.","lead":"This paper combines three existing building blocks (wavelet-based convolution, a scale-feature pyramid, and dynamic convolution) into a YOLOv8-style oriented object detector for aerial images. It reports slightly better accuracy and fewer parameters than a recent method on three remote sensing benchmarks, but the paper's tables and text contradict each other.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"UCAS-AOD baseline undermines cross-method comparability: a plain YOLOv8s scores 95.83 while all cited methods are ~89, implying a split or metric mismatch and making the claimed SOTA gains unsupported.","rationale":"The paper's central claim is empirical: after adding ASFP, C2f-WTC, and C2f-GDC to YOLOv8s, the model reaches or beats SOTA mAP on DOTAv1.0, DIOR-R, and UCAS-AOD while using 21.6M parameters and fewer FLOPs. For that claim to be true, the mAP column must be measured under the same protocol as the cited baselines: same train/test split, same rotated-box format, same IoU or matching rule, same input size, and same evaluation set. The paper provides no code, so the tables are the only evidence. Table V is the most damaging piece of in-scope evidence: the untouched YOLOv8s baseline reaches 95.83 on UCAS-AOD, while every external method in Table III is between 89.02 and 89.99. A 6-point baseline gap cannot be explained by the proposed modules, which add only +2.03 points. It points to a different evaluation setting. Table I's per-class row compounds the problem: the values attributed in the prose to SV and ST are, in the table, the ST and SBF entries respectively, so even the mapping of numbers to categories is not stable. Table V also contradicts the prose: ASFP is claimed to add 0.52 points but the table shows +1.20, and the sentence crediting C2f-WTC with 97.84 actually describes the S+W row. These inconsistencies do not by themselves prove the method is bad, but they mean the central claim is not supported by the published numbers. The concrete test of re-running a plain YOLOv8s baseline under the official oriented evaluation protocol on UCAS-AOD is decisive: it either reproduces approximately 95.83, which would require every cited baseline to be rerun, or it does not, in which case the claimed SOTA comparison is void. The reader's REJECT verdict is appropriate; my read does not move it.","tokens_in":12234,"tokens_out":6691,"duration_ms":57111,"concrete_test":"Re-run the vanilla YOLOv8s baseline, without ASFP/WTC/GDC, on UCAS-AOD using the exact oriented-evaluation code and train/test split used by the cited methods (e.g., S2A-Net's setup), with the paper's stated 640x640 input and 100-epoch schedule, and report per-class AP for Car and Airplane plus mAP50. If the result is not approximately 95.83, then the Table III comparison is invalid and the claimed UCAS-AOD superiority is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in the abstract and Section III-C, is that the proposed YOLOv8s variant beats or matches SOTA oriented detectors at lower parameter and FLOP cost. The load-bearing condition is that Tables I-III compare numbers produced under the same protocol as the cited baselines. That condition fails on the paper's own evidence. In Section III-D, Table V reports the unmodified YOLOv8s baseline as 95.83 mAP50 on UCAS-AOD, while Table III lists the best prior method as 89.99. A 5.84-point advantage for a vanilla baseline over all listed methods cannot be explained by the proposed modules, which add only 2.03 points; it implies a different train/test split, a different label convention (e.g., horizontal rather than rotated boxes), a different IoU or matching rule, or a different test subset. The same protocol ambiguity affects DOTAv1.0 and DIOR-R. The tables also contradict their own prose: the sentence citing SV=70.29 and ST=75.60 in Table I does not match the columns (SV=82.90, ST=70.29, SBF=75.60), and the ablation text says ASFP adds 0.52 points while Table V shows +1.20, and credits C2f-WTC with 97.84, which is actually the S+W row. Because every headline comparison depends on these numbers, the central claim is not presently established; if protocols were matched, the reported margins might vanish.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a YOLOv8-based oriented object detector for aerial imagery, combining three modules: ASFP (an adaptive scale feature pyramid that merges P2/P3 features through OKM-CSP), C2f-WTC (a C2f variant using Haar wavelet transforms), and C2f-GDC (a C2f variant using ghost dynamic convolutions). The authors report mAP values of 78.32 on DOTAv1.0, 67.32 on DIOR-R, and 97.86 on UCAS-AOD with 21.6M parameters, and claim that the model matches or exceeds state-of-the-art oriented detectors while being more lightweight. The paper includes architecture figures, equations, comparison tables, and an ablation study on UCAS-AOD.","tokens_in":12555,"tokens_out":6485,"duration_ms":50178,"significance":"If the empirical claims were reliable, the paper would offer a practically useful lightweight oriented detector, and the module designs are clearly described. The reporting of parameter and FLOP counts in Tables II and IV is a useful feature. However, the central comparison is undermined by internal inconsistencies and by an apparent protocol mismatch: the vanilla YOLOv8s baseline already exceeds all cited UCAS-AOD methods by a large margin, and the text contradicts its own tables. As a result, the claimed state-of-the-art performance is not currently established.","major_comments":[{"comment":"The UCAS-AOD results imply a protocol mismatch that invalidates the headline comparison. Table V reports the unmodified YOLOv8s baseline at 95.83 mAP50, while Table III lists the best prior method (S2A-Net) at 89.99. The proposed modules add at most 2.03 points over the baseline (97.86 vs 95.83), so the 5.84-point gap between a vanilla baseline and all cited methods cannot be explained by the proposed architecture. This gap points to different training splits, input resolutions, label conventions (rotated vs horizontal boxes), IoU thresholds, or test subsets. Because the protocols of the cited methods are not specified and the numbers are evidently not comparable, the claimed superiority of 97.86 vs 89.99 is unsupported. The same concern extends to the DOTAv1.0 and DIOR-R comparisons in Tables I and II, where cited results are taken from their original papers.","section":"Section III.C, Tables III and V"},{"comment":"The text states that for small vehicles (SV) the model achieves 70.29 AP and for storage tanks (ST) 75.60 AP. With the column order stated in Table I (PL, BD, BR, GTF, SV, LV, SH, TC, BC, ST, SBF, RA, HA, SP, HC), the Ours row gives SV=82.90, ST=70.29, and SBF=75.60. The prose values are therefore assigned to the wrong categories, and the table itself must be re-verified.","section":"Section III.C vs Table I"},{"comment":"The ablation narrative does not match the table. The text says ASFP increases mAP by 0.52%, but the S row gives 97.03 vs the baseline 95.83, a 1.20-point gain. The text credits C2f-WTC with 97.84, but the W-only row is 96.94; 97.84 is the S+W row. These mismatches mean the ablation study, which is the main evidence for the individual contributions, is not internally consistent.","section":"Section III.D vs Table V"},{"comment":"The reported computational cost for the proposed model is inconsistent. Table II lists Ours as 21.6M parameters and 52.8 G FLOPs at 800x800 input, while Table IV lists Ours as 21.6M/11.4M parameters and 93.0/76.3 G FLOPs. If both entries describe the same YOLOv8s-based model, the discrepancy in FLOPs needs clarification; if they describe different configurations, the efficiency comparison in Table II is not the one claimed in the text.","section":"Section III.C, Tables II and IV"}],"minor_comments":[{"comment":"The text states that the okbranch and identity have channel ratios e and (1-e), but the equations define x'=e/(1+e)x and x''=1/(1+e)x. For e=0.25, these give 0.2/0.8 rather than 0.25/0.75. Please align the equations with the stated split.","section":"Section II.A, Eqs. (1)-(2)"},{"comment":"The formula for C2f-GDC uses Ydynamic1 but never uses Ydynamic2; please correct the notation or the equation.","section":"Section II.B, Eq. (7)"},{"comment":"References [17] and [18] cite the same paper; the duplicate should be removed.","section":"References"},{"comment":"The author list contains '4th Yi Xiao' followed immediately by '7th XianChuan Yu', which appears to omit the 5th and 6th authors; please fix the numbering.","section":"Author list"},{"comment":"The paper states a 5:2:3 split into 755/302/452 images, which sums to 1509; please reconcile with the dataset size and standard split.","section":"Section III.A"}],"recommendation":"reject","confidential_remarks":"The manuscript has several internal contradictions that cannot be resolved by minor edits. The most serious is that the baseline YOLOv8s already scores far above all cited methods on UCAS-AOD, indicating incomparable evaluation protocols; unless the authors re-run all comparisons under a single protocol and correct the numerical inconsistencies, the headline claims are unsupported. I recommend rejection, with the possibility of resubmission after a full empirical re-evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper combines several published building blocks—SPDConv, Omni-Kernel, wavelet convolution, Ghost dynamic convolution—into a YOLOv8s oriented detector, and claims SOTA on DOTA, DIOR-R, and UCAS-AOD. The architecture is plausible, and the parameter count (21.6M) is genuinely lower than DecoupleNet's 23.3M. If the numbers were clean, this would be a modest but useful incremental contribution.\n\nThey are not. The decisive problem is the UCAS-AOD experiment. In the paper's own ablation (Table V), the unmodified YOLOv8s baseline scores 95.83 mAP50, while all prior methods in Table III are around 89. A vanilla baseline beating every compared method by six points cannot be explained by the proposed modules, which add only about two points. The likely cause is a protocol mismatch: the paper divides UCAS-AOD into 5:2:3 train/val/test, which is not the protocol used by the compared papers. That makes the headline 97.86 mAP unsupported, and the same concern hovers over the DOTA and DIOR-R comparisons.\n\nThe internal inconsistencies compound this. The text in III.C assigns SV=70.29 and ST=75.60, but Table I shows those values under ST and SBF. The ablation text says ASFP adds 0.52 points; the table shows +1.20, and it credits 97.84 to C2f-WTC when that row also includes the S module. Equations (1) and (2) describe a channel split but actually define a scaling—with e=0.25, x' gets 20% of the channels, not 25%. The ablation parameter counts don't add up: S+W alone is listed as 23.29M, but all three modules together are 21.63M. No code is released, and the paper omits citations to the original WTConv and dynamic-convolution works, which is sloppy but minor compared with the empirical issues.\n\nThe best part of the paper is the design: pulling P2 detail through SPDConv to get small-object features without the usual FLOP overhead, and the wavelet/dynamic-convolution modules are sensible components. But the evaluation is broken. This needs a full rework with matched protocols, corrected tables, and a coherent ablation before it can be taken seriously. For the reading group, it's useful only as an example of how protocol mismatches can invalidate a comparison. I would not cite it, and I would desk reject it in its current form.","headline":"A sensible module combination, but the experimental evidence is internally contradictory and the UCAS-AOD comparison is not credible.","tokens_in":13076,"tokens_out":8082,"would_cite":false,"duration_ms":64185,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that adding a wavelet-based feature pyramid and dynamic Ghost convolution to YOLOv8s yields an oriented detector that matches or beats state-of-the-art rotated detectors on aerial benchmarks while using fewer parameters…","keywords":["oriented object detection","rotated bounding box","aerial images","small object detection","YOLOv8","lightweight network","wavelet transform","feature pyramid"],"falsifier":"Re-run the proposed model and the compared methods under one identical protocol: the same split of UCAS-AOD, the same input resolution, and the same mAP metric, and check whether the 95.83 mAP YOLOv8s baseline and the 97.86 mAP final result persist. If the baseline drops to the published 89-90 range under the standard protocol, the claimed small-object advantage is a protocol artifact rather than a property of the modules.","tokens_in":12050,"feed_emoji":"🛰️","tokens_out":4164,"duration_ms":37263,"temperature":0.7,"pith_summary":"This paper proposes a rotated-bounding-box detector for aerial images, built on YOLOv8s, aimed at detecting small objects without a heavy computational footprint. It introduces three components: an Adaptive Scale Feature Pyramid (ASFP) that reuses P2-layer detail cheaply, a wavelet-transform C2f module (C2f-WTC) that enlarges receptive fields with logarithmic parameter growth, and a C2f-Ghost Dynamic Convolution module (C2f-GDC) that keeps parameters low. The reported results are 78.32 mAP on DOTAv1.0, 67.32 mAP on DIOR-R, and 97.86 mAP on UCAS-AOD with 21.6M parameters, competitive with or better than larger models such as DecoupleNet and S2A-Net. The paper's point is that small-object sensitivity and efficiency can be combined in a single-stage oriented detector.","feed_headline":"21.6M-parameter YOLOv8 beats heavier oriented detectors","feed_subtitle":"Wavelet-based feature pyramid lifts small-object accuracy on DOTA, DIOR-R, and UCAS-AOD at lower compute.","key_machinery":"The central mechanism is a set of three modules inserted into YOLOv8s. ASFP distills small-object-rich features from the P2 layer using SPDConv and merges them into the P3 detection path, avoiding the computational blow-up of a dedicated P2 detection head. C2f-WTC applies a recursive Haar wavelet transform to split features into low- and high-frequency sub-bands, runs small-kernel convolutions per sub-band, and recombines with the inverse wavelet transform, yielding larger effective receptive fields at logarithmic parameter cost. C2f-GDC replaces bottleneck convolutions with dynamic expert convolutions whose added parameters scale with the number of experts, not kernel size. The OKM-CSP module with global, large, and local branches carries multi-scale feature extraction inside the neck.","core_discovery":"The paper claims that a YOLOv8s backbone augmented with ASFP, C2f-WTC, and C2f-GDC can outperform or match heavier oriented detectors while using fewer parameters. On DOTAv1.0 it reports 78.32 mAP against DecoupleNet's 78.04; on DIOR-R it reports 67.32 mAP with 21.6M parameters and 52.8 GFLOPs against DecoupleNet's 23.3M parameters and 92.3 GFLOPs; on UCAS-AOD it reports 97.86 mAP against S2A-Net's 89.99. The central assertion is that small-object detail can be recovered from the P2 layer without the usual cost of a full extra detection layer, and that wavelet-decomposed convolutions give multi-scale receptive fields whose parameter count grows logarithmically rather than quadratically with receptive field size.","pith_inferences":["If the reported comparisons assume identical evaluation protocols, then the wavelet-plus-P2 recipe could plausibly transfer to other single-stage detectors or to segmentation backbones, since it targets low-level detail without a full high-resolution detection head.","A direct testable extension is to vary input resolution on DOTAv1.0: if ASFP's small-object gain is real, it should persist or grow at lower resolutions where P2 detail matters most.","Because the paper's own YOLOv8s baseline already scores 95.83 mAP on UCAS-AOD while every published competitor sits near 89, the 97.86 result is meaningful only if the training split, input size, and metric match those of the compared methods.","The dynamic expert convolution in C2f-GDC is architecture-agnostic, so its parameter savings could be measured on other backbones to see whether the efficiency gain is specific to YOLOv8s."],"forward_implications":["On DOTAv1.0, the method reports 78.32 mAP, surpassing S2A-Net's 76.11 with clear gains on small vehicles (70.29 AP) and storage tanks (75.60 AP).","On DIOR-R, it reports 67.32 mAP with 21.6M parameters and 52.8 GFLOPs, beating DecoupleNet's 67.08 mAP at lower computational cost.","On UCAS-AOD, it reports 97.86 mAP, with per-class AP of 96.32 for cars and 99.21 for airplanes.","Ablation on UCAS-AOD attributes a 0.52-point gain to ASFP alone and shows the combined modules reach 97.86 mAP against a YOLOv8s baseline of 95.83, at an added parameter cost of about 10M.","The parameter count of 21.6M with the reported FLOPs suggests the architecture is positioned for deployment on resource-constrained platforms such as drones or edge devices."],"supporting_citations":[{"why":"DecoupleNet is the main efficiency-and-accuracy baseline the proposed model claims to beat on DIOR-R and DOTAv1.0.","marker":"[19]"},{"why":"S2A-Net is the primary accuracy baseline on DOTAv1.0 and UCAS-AOD, with per-class AP numbers the paper compares against.","marker":"[27]"},{"why":"DOTAv1.0 is the main benchmark dataset, providing the 15-category evaluation used for the headline mAP.","marker":"[23]"},{"why":"DIOR-R is the second dataset used to evaluate small-object and multi-scale detection performance.","marker":"[24]"},{"why":"UCAS-AOD is the small-object-focused dataset used for the main ablation study and the reported 97.86 mAP.","marker":"[25]"},{"why":"SPDConv supplies the mechanism ASFP uses to distill small-object features from the P2 layer without strided convolutions.","marker":"[20]"},{"why":"The Omni-Kernel Module is the basis for the OKM-CSP module's global, large, and local branches.","marker":"[21]"},{"why":"GhostModule (dynamic convolution with expert networks) is the basis for the C2f-GDC module's lightweight design.","marker":"[22]"}],"fun_headline_variants":["Lighter YOLOv8 for oriented objects: less compute, better small-object mAP","21.6M-param YOLOv8 outperforms heavier oriented detectors","Wavelet-based YOLOv8: small-object oriented detection, lower FLOPs","Oriented object detection with wavelet feature pyramid: fewer params, similar mAP","YOLOv8 with ASFP and wavelet: lighter than DecoupleNet, better on DOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline accuracy numbers assume the same training splits, input resolutions, and evaluation conventions as the published numbers they are compared with, so the reported gains over competitors stand only if those protocols match.","fun_headline_variants_meta":{"raw":{"variants":["Lighter YOLOv8 for oriented objects: less compute, better small-object mAP","21.6M-param YOLOv8 outperforms heavier oriented detectors","Wavelet-based YOLOv8: small-object oriented detection, lower FLOPs","Oriented object detection with wavelet feature pyramid: fewer params, similar mAP","YOLOv8 with ASFP and wavelet: lighter than DecoupleNet, better on DOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001785,"raw_usage":{"total_tokens":7053,"prompt_tokens":979,"completion_tokens":6074,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":5963}},"tokens_in":595,"tokens_out":6074,"duration_ms":42859,"temperature":1.0,"reasoning_tokens":5963,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:56:59.936926+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the proposed model and the compared methods under one identical protocol: the same split of UCAS-AOD, the same input resolution, and the same mAP metric, and check whether the 95.83 mAP YOLOv8s baseline and the 97.86 mAP final result persist. If the baseline drops to the published 89-90 range under the standard protocol, the claimed small-object advantage is a protocol artifact rather than a property of the modules.","supporting_citations":[{"cited_title":"Decou- plenet: A lightweight backbone network with efficient feature decoupling for remote sensing visual tasks","cited_arxiv_id":null,"evidence_quote":"DecoupleNet is the main efficiency-and-accuracy baseline the proposed model claims to beat on DIOR-R and DOTAv1.0."},{"cited_title":"Align deep features for oriented object detection","cited_arxiv_id":null,"evidence_quote":"S2A-Net is the primary accuracy baseline on DOTAv1.0 and UCAS-AOD, with per-class AP numbers the paper compares against."},{"cited_title":"Dota: A large- scale dataset for object detection in aerial images","cited_arxiv_id":null,"evidence_quote":"DOTAv1.0 is the main benchmark dataset, providing the 15-category evaluation used for the headline mAP."},{"cited_title":"Anchor-free oriented proposal generator for object detection","cited_arxiv_id":null,"evidence_quote":"DIOR-R is the second dataset used to evaluate small-object and multi-scale detection performance."},{"cited_title":"Orientation robust object detection in aerial images using deep convolutional neural network","cited_arxiv_id":null,"evidence_quote":"UCAS-AOD is the small-object-focused dataset used for the main ablation study and the reported 97.86 mAP."},{"cited_title":"No more strided convolutions or pooling: A new cnn building block for low-resolution images and small objects","cited_arxiv_id":null,"evidence_quote":"SPDConv supplies the mechanism ASFP uses to distill small-object features from the P2 layer without strided convolutions."},{"cited_title":"Omni-kernel network for image restoration","cited_arxiv_id":null,"evidence_quote":"The Omni-Kernel Module is the basis for the OKM-CSP module's global, large, and local branches."},{"cited_title":"Parameternet: Parameters are all you need for large-scale visual pretraining of mobile networks","cited_arxiv_id":null,"evidence_quote":"GhostModule (dynamic convolution with expert networks) is the basis for the C2f-GDC module's lightweight design."}],"review_version":1}