{"id":"a5def5f5-50db-4b48-9399-aea3afb77b48","arxiv_id":"2501.01855","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A real-time end-to-end UAV object detector that combines spatial and frequency-domain features, gaining 3.1 AP over RT-DETR on VisDrone.","lead":"UAV-DETR adapts the real-time detection transformer RT-DETR to drone imagery by adding frequency-domain feature fusion, frequency-focused downsampling, and a semantic alignment module. On the VisDrone benchmark it raises average precision by about 3 points over the RT-DETR baseline while keeping real-time speeds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (3)'s frequency-domain product is dimensionally invalid, leaving the key FF module unspecified and the reported gains unattributable to the described method.","rationale":"The reader's weakest assumption concerns baseline training comparability, which is a valid reproducibility concern. My stress test identifies an additional, more fundamental internal issue: the mathematical definition of the key FF module is ambiguous. Equation (3) appears to multiply a complex Fourier-transformed feature by a real spatial-domain feature, which is not a defined tensor operation. This is not merely a missing baseline detail; it directly affects whether the proposed architecture is even implementable as described. If the formula is a typo and the actual operation is a learned filter in the frequency domain, the authors should provide the corrected equation and implementation. The paper does have supportive ablations and a parameter study for Inner-SIoU, and the gains are plausible, but the ill-defined FF operation prevents full verification. Since the reader already assigned CONDITIONAL, my concern does not change the verdict; it strengthens the conditions: the authors must clarify Eq. (3) and release code. I therefore set verdict_should_be to UNCHANGED (still CONDITIONAL). I partially agree with the reader because we both flag reproducibility, but I focus on an internal inconsistency rather than training-protocol comparability.","tokens_in":10426,"tokens_out":5287,"duration_ms":54503,"concrete_test":"Obtain the official code from the project page (or ask the authors) and execute a forward pass of the FF module on a real input, checking whether Eq. (3)'s multiplication aligns a complex frequency-domain tensor with a real spatial-domain tensor; if the shapes do not match, the equation is invalid as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—the 3.1 AP gain over RT-DETR-R18 on VisDrone (Table I)—is attributed to three modules whose shared core is the Frequency-Focused (FF) module (Eq. 3). In Section III-A, x_F = α · IF(F(Conv1×1(x_sc)) · Conv1×1(x_sc)) + β · x_sc. Here F(Conv1×1(x_sc)) is a complex-valued frequency-domain tensor, while Conv1×1(x_sc) is a real-valued spatial-domain tensor. Element-wise multiplication between these is undefined without a transform matching; a standard 1×1 convolution on a real spatial tensor cannot be multiplied with the complex Fourier coefficients. If implemented literally, the operation fails; if implemented as a learned frequency-domain filter (as the text suggests), the formula is incorrect and the actual operation is not specified. Because FF appears inside MSFF-FE, FD, and SAC, the entire architecture's behavior in the experiments is ambiguous. The paper also does not release code or provide a precise pseudocode for this operation, so readers cannot disambiguate. The ablation in Table III only adds modules cumulatively and does not isolate the FF operation from the concurrent increase in GFLOPs (60 to 77 for R18), so the empirical gains in Table I may not reflect the described frequency mechanism. This is an internal inconsistency that should be resolved before the headline improvement can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes UAV-DETR, an end-to-end DETR-style object detector for UAV imagery, built on RT-DETR. It introduces three modules: a multi-scale feature fusion with frequency enhancement module (MSFF-FE), a frequency-focused downsampling module (FD), and a semantic alignment and calibration module (SAC), plus an Inner-SIoU loss. The central claim is a 3.1 AP and 4.2 AP50 improvement over the RT-DETR-R18 baseline on the VisDrone validation set, with similar gains on UAVVaste, while retaining real-time inference. The paper reports ablations, speed measurements, and heatmap visualizations to support the design.","tokens_in":10711,"tokens_out":4875,"duration_ms":49262,"significance":"If validated, the result is useful for UAV object detection: it demonstrates that an end-to-end, NMS-free detector can be competitive on aerial imagery, where small, occluded objects are common, and that frequency-domain processing can be integrated into a real-time DETR. The paper's main strengths are its clear architectural motivation, the inclusion of ablation experiments on VisDrone, generalization experiments on a second dataset (UAVVaste), and speed measurements showing that real-time operation is roughly maintained. The claimed gains are nontrivial (3.1 AP over RT-DETR-R18), and the work addresses an application-relevant problem. However, the central claim is not yet fully substantiated because of ambiguities in the frequency-domain formulation, incomplete baseline comparability, and the absence of statistical evidence such as error bars or repeated runs.","major_comments":[{"comment":"The Frequency-Focused (FF) module, as written, mixes spatial and frequency domains in a way that is not adequately specified. In Eq. (3), F(Conv1x1(x_sc)) is a complex-valued frequency-domain tensor while Conv1x1(x_sc) is a real-valued spatial-domain tensor; the element-wise product is mathematically defined as a complex-by-real scaling, but it is not a frequency-domain filtering operation with learned coefficients, because the multiplier is derived from spatial coordinates and cannot adjust the phase of the frequency components. If the intended operation is a learned filter in the frequency domain, the equation does not describe how the filter is parameterized. Since FF is used inside MSFF-FE, FD, and SAC, the reported experimental gains cannot be attributed to the described frequency mechanism without a precise definition or pseudocode. Please clarify the exact tensor operations, including shapes and complex arithmetic, or revise the description to match the implementation.","section":"III-A, Eq. (3)"},{"comment":"The comparability of the baseline results is not established. The paper gives the full training recipe for UAV-DETR (400 epochs, batch size 4, 640x640, Mosaic probability 1, mixup probability 0.2), but it does not state that RT-DETR-R18, RT-DETR-R50, or the other competitors in Table I were retrained under this identical protocol; the table likely reports numbers from the respective original papers. Because the central claim is a 3.1 AP improvement over RT-DETR-R18, the comparison is only meaningful if the baseline is trained with the same augmentations, input size, and epoch budget. Please report the training setup of all compared models or, ideally, retrain the relevant baselines under the same conditions and provide those numbers.","section":"IV-A and Table I"},{"comment":"The ablation study does not yet support the claim that each module contributes independently. The MSFF-FE addition increases GFLOPs from 60 to 77 (about 28%), so the AP gain from 27.1 to 28.4 may be partly a capacity effect rather than a frequency-enhancement effect. The FD module shows no AP improvement (28.4 to 28.4) and only a 0.2 AP50 improvement. Furthermore, all results appear to be from a single run, with no error bars or repeated-seed variance, and the Inner-SIoU ratio of 1.25 is selected on the same validation set (Table IV). At minimum, please provide variance over multiple runs and a matched-compute comparison, and discuss the FD result explicitly.","section":"Table III and Table IV"},{"comment":"The Inner-SIoU loss is not precisely defined. Equation (10) reads L_Inner-SIoU = L_SIoU + IoU - Inner-SIoU, but it is unclear whether 'IoU' denotes the standard IoU loss (1 - IoU) or the IoU metric, and whether 'Inner-SIoU' denotes an IoU value or a loss term. This makes the loss formula ambiguous and difficult to reproduce. Please define all terms and, if this is a combination of Inner-IoU and SIoU losses from the cited works, give the exact functional form used in training.","section":"III-D, Eq. (10)"}],"minor_comments":[{"comment":"The GFLOPs values for RT-DETR-R18 and RT-DETR-R50 differ between Table V (60 and 130) and Tables I and II (60 and 136/57.3 and 129.9); please make these numbers consistent across tables.","section":"Table V"},{"comment":"The implementation details list 'momentum of 0.9' for AdamW, but AdamW typically uses betas; please clarify whether this is the first momentum parameter or a different optimizer setting.","section":"IV-A"},{"comment":"There is a grammatical error in the introduction: 'align features that from different feature fusion paths' should be 'align features from different feature fusion paths.'","section":"I"},{"comment":"The capitalization of 'Inner-SIoU' is inconsistent (also 'Inner-SIOU'), and the abbreviation should be defined once and used consistently thereafter.","section":"III-D"},{"comment":"The reference for YOLOv9 is cited as 'ECCV2025' in Table I but the bibliography entry [30] is an arXiv preprint; please correct the venue information.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The main risk to the paper's central claim is not statistical fraud or circularity, but rather that the frequency-domain module is not specified tightly enough for the experimental results to be attributed to the described mechanism. The authors should be asked to provide either code or a precise pseudocode for Eq. (3), to clarify the baseline training protocol, and to report variance. If these issues are resolved, the paper could be a reasonable contribution to the UAV detection literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid engineering contribution: it takes RT-DETR and adds three modules (MSFF-FE, FD, SAC) plus Inner-SIoU, and reports a 3.1 AP / 4.2 AP50 improvement on VisDrone with similar gains on UAVVaste, while keeping real-time speeds. The ablations show every component helps, and the writing is clear. That is worth taking seriously.\n\nWhat is actually new is the specific combination: frequency-domain processing embedded in multi-scale fusion, a frequency-focused downsampling path, and a semantic alignment module with learned offsets. The building blocks are known (FFT, gating, deformable sampling, Inner-SIoU), but the assembly is original and the empirical result is non-trivial. The authors also honestly flag a failure mode (focusing on irrelevant regions) and do not oversell.\n\nThe soft spots are real but not fatal. First, Eq. (3) is dimensionally confusing: F(Conv1x1(x_sc)) is a complex frequency-domain tensor, and multiplying it by Conv1x1(x_sc) in the spatial domain is not a standard filtering operation. You can make the shapes work in PyTorch, but the semantic mapping between frequency and spatial coordinates is missing. If the intended operation is a learned frequency-domain filter, the formula is wrong; if it is literal, the mechanism is not what the text claims. The paper links a project page but does not include code or pseudocode, so readers cannot disambiguate. Second, there are no error bars, and the baseline comparisons are not fully controlled — the paper gives its own training recipe but does not state that all competitors were retrained identically. Third, the Inner-SIoU ratio (1.25) is tuned on the same validation set used for reporting. These are common issues in applied CV papers, but they limit how much weight the exact numbers can carry.\n\nThe central empirical claim — that the full model improves over RT-DETR — is supported by the ablation, but the attribution to the frequency mechanism is weaker than the abstract implies. The ablation adds modules cumulatively alongside GFLOPs increases, so the gain could partly come from extra capacity.\n\nWho is this for? Researchers working on real-time UAV detection who want a strong baseline and a reasonable bag of tricks. It is not a conceptual breakthrough, but it is a useful, honest empirical study.\n\nShould a serious editor send it to peer review? Yes. The issue with Eq. (3) is fixable in revision, and the empirical results are worth checking. I would recommend major revision: clarify the FF formulation, release code, add error bars, and control baselines more carefully.","headline":"A competent UAV detection architecture with a real 3.1 AP gain over RT-DETR, but the key frequency-focused module is underspecified in Eq. (3), so the reported gains are not yet attributable to the described mechanism.","tokens_in":11218,"tokens_out":2260,"would_cite":true,"duration_ms":24589,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Frequency-domain fusion lifts drone detection by 3.1 AP","keywords":["object detection","unmanned aerial vehicle","DETR","frequency domain","feature fusion","small object detection","real-time detection","VisDrone"],"falsifier":"Retrain RT-DETR-R18 with the exact UAV-DETR training recipe (400 epochs, batch size 4, input 640x640, Mosaic probability 1, mixup 0.2) and compare AP on the VisDrone validation set, or verify that ablating the frequency branch specifically degrades AP on small objects; if the gap between baseline and UAV-DETR collapses or the frequency terms have no effect, the central claim is unsupported.","tokens_in":10264,"feed_emoji":"🚁","tokens_out":6195,"duration_ms":49182,"temperature":0.7,"pith_summary":"This paper sets out to establish that a detection transformer (DETR, an end-to-end detector that predicts objects without non-maximum suppression or hand-designed anchors) can be made effective for drone imagery, where objects are small and often occluded, by deliberately preserving frequency-domain information during feature fusion and downsampling and by aligning features from different fusion paths. The proposed UAV-DETR builds on the real-time RT-DETR detector and adds three modules: multi-scale feature fusion with frequency enhancement, frequency-focused downsampling, and semantic alignment and calibration, together with an Inner-SIoU bounding-box loss. On the VisDrone dataset, the ResNet-18 variant raises AP by 3.1 percentage points and AP50 by 4.2 points over the RT-DETR-R18 baseline, with similar gains on the UAVVaste dataset, while keeping real-time inference speeds. If these results hold, they suggest that frequency-domain cues are a practical lever for small-object detection in aerial imagery without sacrificing the deployment advantages of end-to-end detectors.","feed_headline":"Frequency-domain fusion lifts drone detection by 3.1 AP","feed_subtitle":"Adds frequency-aware fusion to a real-time DETR, raising small-object detection on drone imagery.","key_machinery":"The argument is carried by three modules plus a loss. MSFF-FE (multi-scale feature fusion with frequency enhancement) splits features along channels, applies a Focus module and Fourier-domain filtering, then recombines spatial and frequency branches through learned weights; FD (frequency-focused downsampling) processes one channel-split branch with a frequency-focused filter and the other with strided convolutions and max-pooling before re-concatenation; SAC (semantic alignment and calibration) upsamples one feature, applies the frequency-focused filter, and uses learned offsets with grid sampling to align the two features before a weighted sum. Inner-SIoU loss replaces GIoU by combining the angle, distance, and shape penalties of SIoU with an Inner-IoU computed on boxes scaled by 1.25, which the authors say improves sensitivity for small objects.","core_discovery":"The central claim is that existing end-to-end detectors, designed for natural images, underperform on UAV imagery because they lose high-frequency details during feature fusion and downsampling, and because features from different fusion paths are not spatially aligned. UAV-DETR addresses this with three components: MSFF-FE fuses multi-scale features in both spatial and frequency domains, using Fourier transforms and a learned gating mechanism; FD applies frequency-focused processing while downsampling to retain critical spatial details; and SAC learns 2D offsets to geometrically align features from different paths before a weighted summation. Together with Inner-SIoU loss, the method reports AP of 29.8 and AP50 of 48.8 on VisDrone for the ResNet-18 variant, versus 26.7 and 44.6 for the RT-DETR-R18 baseline, and also outperforms comparable-cost detectors on UAVVaste. The authors interpret these results as evidence that frequency-domain information and cross-path alignment are effective, generalizable mechanisms for UAV object detection.","pith_inferences":["If the frequency-enhancement modules are the real source of the gain, a similar recipe should transfer to other small-object domains such as satellite imagery, medical imaging, or long-range autonomous driving, where high-frequency details and occlusion dominate.","The learned-offset alignment in SAC could be reused outside detection, for example in multi-scale feature fusion for segmentation or in fusing features from different sensor modalities with spatial misalignment.","A direct test of the paper's claim would be to retrain the RT-DETR baselines under the paper's exact training protocol (400 epochs, batch size 4, input 640x640, Mosaic probability 1, mixup 0.2); if the 3.1 AP gap mostly disappears, the modules' contribution is smaller than reported.","The FPS comparison is measured in 32-bit floating-point precision on a single GPU; on embedded UAV hardware the relative speed cost of the added modules could differ, so real-time deployment claims should be re-verified on target hardware."],"forward_implications":["End-to-end detection without NMS or anchor tuning is viable for UAV imagery while matching or beating manual-pipeline detectors at similar computational cost.","Preserving high-frequency components during fusion and downsampling is a transferable strategy, with gains appearing on both VisDrone and the smaller UAVVaste dataset.","The three modules compose additively: ablation shows AP rising from 26.7 to 29.8 as Inner-SIoU, MSFF-FE, FD, and SAC are added in sequence.","Real-time performance is largely retained: UAV-DETR-R18 runs at 124 FPS versus 183 FPS for RT-DETR-R18, and the R50 variant at 65 FPS versus 89 FPS, measured in 32-bit floating-point precision.","Frequency-focused operations help small and occluded objects specifically, as indicated by heatmaps showing stronger attention on small objects and their surroundings."],"supporting_citations":[{"why":"Supplies the RT-DETR baseline architecture and training recipe that UAV-DETR builds on and compares against.","marker":"[6]"},{"why":"Introduces the end-to-end DETR paradigm that eliminates NMS and anchor boxes.","marker":"[4]"},{"why":"Defines Inner-IoU with auxiliary scaled boxes used in the Inner-SIoU loss.","marker":"[19]"},{"why":"Defines SIoU whose angle, distance, and shape penalties are combined in the new loss.","marker":"[18]"},{"why":"Supplies the Focus module used to compress spatial information in MSFF-FE.","marker":"[20]"},{"why":"Provides the cross-stage partial strategy that partitions features in MSFF-FE.","marker":"[21]"},{"why":"Supplies the differentiable grid sampling used by SAC to align features.","marker":"[23]"},{"why":"Provides the VisDrone dataset used for the main experiments.","marker":"[24]"},{"why":"Provides the UAVVaste dataset used to test generalization.","marker":"[25]"}],"fun_headline_variants":["UAV-DETR: frequency fusion adds 3.1 AP for drone imagery","Drone DETR improves AP 3.1 via cross-scale frequency fusion","Frequency-aware detection transformer lifts drone AP by 3.1","UAV-DETR retains spatial details, gains 3.1 AP on VisDrone","Efficient DETR for drones uses frequency domain to gain 3.1 AP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported improvements assume the RT-DETR baselines and other compared detectors were evaluated under training conditions comparable to UAV-DETR's; the paper details its own recipe (400 epochs, batch size 4, input 640x640, Mosaic probability 1, mixup 0.2) but does not state that all baselines were retrained with the same protocol, so part of the 3.1 AP gap could reflect training differences rather than the new modules.","fun_headline_variants_meta":{"raw":{"variants":["UAV-DETR: frequency fusion adds 3.1 AP for drone imagery","Drone DETR improves AP 3.1 via cross-scale frequency fusion","Frequency-aware detection transformer lifts drone AP by 3.1","UAV-DETR retains spatial details, gains 3.1 AP on VisDrone","Efficient DETR for drones uses frequency domain to gain 3.1 AP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000842,"raw_usage":{"total_tokens":3685,"prompt_tokens":976,"completion_tokens":2709,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":2616}},"tokens_in":592,"tokens_out":2709,"duration_ms":20257,"temperature":1.0,"reasoning_tokens":2616,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:18:23.385151+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain RT-DETR-R18 with the exact UAV-DETR training recipe (400 epochs, batch size 4, input 640x640, Mosaic probability 1, mixup 0.2) and compare AP on the VisDrone validation set, or verify that ablating the frequency branch specifically degrades AP on small objects; if the gap between baseline and UAV-DETR collapses or the frequency terms have no effect, the central claim is unsupported.","supporting_citations":[{"cited_title":"Detrs beat yolos on real-time object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the RT-DETR baseline architecture and training recipe that UAV-DETR builds on and compares against."},{"cited_title":"End-to-end object detection with transformers,","cited_arxiv_id":null,"evidence_quote":"Introduces the end-to-end DETR paradigm that eliminates NMS and anchor boxes."},{"cited_title":"Cspnet: A new backbone that can enhance learning capability of cnn,","cited_arxiv_id":null,"evidence_quote":"Provides the cross-stage partial strategy that partitions features in MSFF-FE."},{"cited_title":"Autonomous, onboard vision-based trash and litter detection in low altitude aerial images collected by an unmanned aerial vehicle,","cited_arxiv_id":null,"evidence_quote":"Provides the UAVVaste dataset used to test generalization."}],"review_version":1}