{"id":"556343b4-985e-47e6-9e9f-3100ce9c75d8","arxiv_id":"2506.13501","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Adding frequency-spatial attention (FSTB) and a training-time corruption branch (HDC) improves detection and segmentation of overlapping objects by 0.3-4.3 AP on four X-ray datasets.","lead":"This paper proposes FOAM, a plug-in framework that adds frequency-domain attention blocks and a corruption-based consistency loss to standard object detectors, reporting accuracy gains on X-ray prohibited item and pneumonia detection tasks. It shows consistent but small improvements across several detectors, with the largest gains on DINO-based models; specialists may find the frequency-domain recipe useful, but the evidence is weakened by missing baselines and error bars.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set hyperparameter selection in HDC (Table VII and Fig. 5) inflates reported AP gains; the central empirical claim needs validation-based selection.","rationale":"The reader's verdict (CONDITIONAL) already flags test-set hyperparameter selection as one of several issues, but the weakest_assumption field identifies the HDC mechanism's theoretical premise instead. I judge the test-set selection to be the single most load-bearing concern because it directly attacks the reliability of the headline numbers: if the corruption and loss hyperparameters are cherry-picked on the test set, the reported improvements—including the 2.9 AP gain on PIXray-det—are the maximum of a small grid search, not unbiased estimates. This can explain gains of the observed magnitude even if FOAM contributes nothing. The HDC theoretical weakness (Theorem I is a trivial inequality about linear responses) is a concern about the paper's explanation, not about whether the method works empirically; similarly, the sign/notation issues in Eqs. 8, 25, and 29 threaten reproducibility but could be typographical. The missing FAPID/FDTNet baselines weaken the 'surpassing specialized detectors' subclaim but do not invalidate the relative gain over DINO. The test-set selection is therefore the one issue that, if confirmed, would require re-evaluating the central empirical claim. Since the reader already assigned CONDITIONAL partly on this basis, my stress-test does not change the verdict; it sharpens the priority: the authors must either use a validation split or report the distribution over hyperparameter choices before the headline numbers can be trusted.","tokens_in":26570,"tokens_out":7121,"duration_ms":74449,"concrete_test":"Reproduce the PIXray-det experiments using a proper train/validation split: hold out a validation subset from the 4046 training images, select the corruption strategy and all hyperparameters (GB sigma, loss type, loss weight, learning rate) using only validation AP, then retrain the selected configuration on the full training set and evaluate on the test set exactly once. Report the final F-DINO and DINO AP with 3–5 random seeds as mean±std. If the gain over DINO falls below 2.9 or becomes statistically insignificant, the central claim needs to be qualified as configuration-dependent rather than a robust improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FOAM consistently improves state-of-the-art detectors, e.g., F-DINO gains 2.9 box AP over DINO on PIXray-det. However, the key components of the method—the corruption strategy and its severity (Table VII), the consistent-loss type, its weight, and the learning rate (Fig. 5)—are selected by maximizing box AP on the PIXray-det test set. Table VII evaluates GB, DU, and GN at 4–5 severity levels each (roughly 14 configurations) and adopts the single best setting (GB, ks=3, sigma=5) as the final model; Fig. 5 then picks the loss type and coefficients using the same test set. This procedure makes the reported AP a maximum over a small grid of configurations, not the performance of a predetermined method. Consequently, the 2.9-point gain over DINO, and the smaller gains on OPIXray, PIXray-seg, and RLP, are optimistically biased and cannot be taken as unbiased estimates of FOAM's true effect. Because the same test set is used for both model selection and final evaluation, there is no protection against overfitting to the test set, and no multiple-comparison correction or validation split is reported. This directly undermines the empirical evidence for the headline claim, independent of whether the HDC mechanism's theoretical motivation or the exact form of Eqs. 8, 25, and 29 is ultimately correct.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FOAM, a plug-in framework for overlapping-object perception. It inserts Frequency Spatial Transformer Blocks (FSTBs) after the backbone to combine spatial- and frequency-domain cues, and adds a training-only Hierarchical De-Corrupting (HDC) branch that aligns features from a corrupted image to the base-branch features via a consistency loss. FOAM is evaluated by adding it to DETR-based, CNN-based, and segmentation detectors on PIXray-det, OPIXray, PIXray-seg, and RLP, reporting gains over baselines with the largest gain being 2.9 AP for F-DINO over DINO on PIXray-det. The paper includes extensive ablations of corruption strategies, loss types, target layers, and attention-map shapes, together with qualitative visualizations of features, sampling points, and predictions.","tokens_in":27009,"tokens_out":9687,"duration_ms":94164,"significance":"If confirmed, FOAM would be a practically useful plug-in: it is architecture-agnostic, improves several modern detectors, and addresses a real need in security X-ray and medical imaging. The breadth of the experiments (four datasets, three tasks, multiple backbones) is a strength, as is the coverage of design choices in the ablations. However, the current evidence is not yet convincing: the headline configuration is selected on the test set, several reported gains are within typical run-to-run variation, the two closest frequency-domain competitors are missing from the comparisons, and the mechanism is not precisely defined in the equations. The contribution is incremental and modular rather than a new task or paradigm; with corrected evaluation and added comparisons it could become a solid applied contribution.","major_comments":[{"comment":"The empirical protocol in Sec. IV-E selects the final model by maximizing box AP on the PIXray-det test set. Table VII evaluates DU, GN, and GB at several severity levels each and adopts the single best setting (GB, ks=3, sigma=5), and Fig. 5 then chooses the consistent-loss type, its coefficient, and the learning rate using the same test set. Because the reported F-DINO result of 67.2 AP is the maximum over this grid, the headline 2.9 AP gain over DINO is not an unbiased estimate of FOAM's effect. Please switch to a validation-based selection procedure (or nested cross-validation) and report only the test-set result of the configuration chosen on the validation set.","section":"Sec. IV-E, Table VII, Fig. 5"},{"comment":"Several of the headline improvements are very small and are reported without variance. For example, Table I reports F-RT-DETR at 61.7 vs. RT-DETR at 61.4 (+0.3 AP), F-DINO(Swin-L) at 73.7 vs. DINO(Swin-L) at 73.3 (+0.4 AP), and F-Mask-R-CNN(X-101) at 65.6 vs. 65.2 (+0.4 AP); Table V reports F-Cascade-R-CNN at 20.2 vs. Cascade-R-CNN at 19.7 (+0.5 AP). With single runs and no error bars, these differences are consistent with training noise under the 12-epoch protocol. Please report means and standard deviations over multiple seeds, or otherwise demonstrate that the differences are statistically stable.","section":"Table I and Table V"},{"comment":"The manuscript motivates FOAM against FAPID [12] and FDTNet [14], stating that their fixed high-pass and SRM filters are limiting, yet neither method appears in the quantitative comparisons in Tables II-IV. Because they are the closest prior frequency-domain approaches for prohibited-item detection, their omission prevents the reader from assessing the central improvement claim over the specialized state of the art. Please include them under the same training and evaluation protocol as the other baselines.","section":"Sec. I, Sec. II-A, Tables II-IV"},{"comment":"The HDC loss is defined inconsistently. Sec. III-B2 states that the loss supervises F^{n+1}_C with F^n_O, but Eq. (8) compares \\hat F^{n-1}_{O,l} with \\hat F^n_{C,l}, and Eq. (9) normalizes by \\sum_i \\exp(F^n_{O,l}(i)) while the numerator is \\exp(F^{n-1}_{O,l}(i)). Additionally, Theorem I analyzes a scalar additive response f+b and shows that reducing b increases the contrast ratio; it does not establish that minimizing the KL loss in Eq. (8) suppresses background response in deep feature maps. Please correct the indexing and either prove the claimed suppression under explicit assumptions or label Theorem I as intuition rather than a formal justification.","section":"Sec. III-B2, Eqs. (8)-(11), Theorem I"},{"comment":"Numbers across tables are not mutually consistent. Table I reports RT-DETR (ResNet-50) at 61.4 APbox, while Table II reports 62.3 for the same detector and backbone; Table VIII's caption says the experiment uses Gaussian Blurring with kernel size 5 and sigma 5, but its best row (l3,l4) gives 67.2, the same value reported for the final model that Table VII selects with ks=3. The reader cannot tell which corruption setting produced which result. Please align the baselines and state the exact settings used in every ablation.","section":"Table I vs. Table II; Table VII vs. Table VIII"}],"minor_comments":[{"comment":"Eq. (29) has unbalanced parentheses and an ambiguous norm placement; please rewrite it so that the operator precedence is clear.","section":"Eq. (29)"},{"comment":"The caption lists four category names for five columns (FO, ST, SC, UT, MU) and omits the definition of SC; please include all five category definitions.","section":"Table III caption"},{"comment":"The caption states that 'KS means the kernel size of Gaussian Noise'; it should say Gaussian Blurring.","section":"Table VII caption"},{"comment":"The boxplot is referenced but the exact grid values for the learning rate and loss coefficient are not reported in the text or caption; please list the tested values.","section":"Fig. 5"},{"comment":"The text says 'FPS decreased by 16 frames'; FPS is already a rate, so please rephrase this as a decrease of 16 frames per second.","section":"Sec. IV-A, Table VI"},{"comment":"The statement that 'the exact form of the consistent loss is not crucial' is hard to reconcile with the ablation in Sec. IV-E3 showing a preference for the Type I loss; please clarify or qualify the statement.","section":"Sec. III-B2"}],"recommendation":"major_revision","confidential_remarks":"The test-set-based model selection in Sec. IV-E is the main methodological obstacle: the reported 67.2 AP is a maximum over a hyperparameter grid evaluated on the test set, so the headline gain is optimistically biased. This is fixable with a validation split, but after correction some of the smaller gains may vanish. The missing comparison with FAPID and FDTNet is also substantive given the paper's positioning. The self-cited baselines (AO-DETR, MMCL, CSPCL) appear to be used as competitors rather than as a way to inflate gains, so I do not see a novelty-disclosure problem there."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The FSTB/HDC combination is a genuinely new assembly of known components, and the gains are consistently positive across detectors and datasets. But the headline 2.9 AP gain on PIXray-det is not an unbiased estimate: the corruption strategy, its severity, the loss type, and the loss weight are all chosen by maximizing AP on the test set (Table VII, Fig. 5). That makes the reported number a maximum over roughly a dozen configurations, not the performance of a predetermined method. The stress-test note lands.\n\nWhat is actually good: FDBA combining magnitude-spectrum self-attention with spatial channel attention is a reasonable idea, and HDC as a training-only corruption-consistency regularizer is clever, even if the theoretical framing is weak. The experiments are broad: multiple detectors, backbones, and four datasets, with ablations that are mostly thorough. The visualization analysis is a nice touch. The method does not change the conceptual foundations of the field, but it is honest incremental progress.\n\nSoft spots, in proportion. The test-set selection is the load-bearing one. Table VII sweeps 14 corruption configurations and picks the best; Fig. 5 then picks loss type and coefficients using the same test set, and the text says each subsequent group uses the best settings from the previous group. That is multiple testing on the test set, no validation split, no correction. The effect on the 2.9 AP gain could easily be a point or more. Second, there are no error bars anywhere; several reported gains are 0.3–0.5 AP, which is within run-to-run noise for DETR-style training. Third, FAPID and FDTNet, the two closest frequency-domain prohibited-item detectors, are missing from the comparisons; that omission undercuts the claim of being a general framework. Fourth, Eqs. 25 and 29 use unclear notation for the modulus operation on complex FFT outputs. Finally, Theorem I is a trivial inequality dressed up as a theorem; it does not constrain the design, and the paper would be better without it.\n\nBottom line: this is a paper for people working on overlapping object detection in X-ray or medical images. The method is plausible, the work is detailed, and the main empirical claim is probably true in direction but not in the reported magnitude. It deserves a serious referee, but the proper outcome is major revision, not acceptance as is. The fixes are clear: redo selection on a validation split, report mean and variance over seeds, include FAPID and FDTNet, and clean up the equations.","headline":"A plausible incremental architecture whose headline AP gains are inflated by test-set hyperparameter selection, but the core idea is worth a serious look after fixes.","tokens_in":27495,"tokens_out":1853,"would_cite":false,"duration_ms":22581,"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":"FOAM boosts DINO by 2.9 AP on overlapping X-ray detection","keywords":["overlapping object perception","frequency domain learning","magnitude spectrum","X-ray prohibited item detection","object detection","instance segmentation","pneumonia detection","consistency loss"],"falsifier":"Measure foreground-background contrast in the feature maps of DINO, DINO+FSTB, and FSTB+HDC under identical training budgets. If box AP rises with HDC but average contour contrast in the feature maps does not increase, then the claimed suppression mechanism is not what drives the gain. Alternatively, replace the corrupted-image branch with an unrelated randomly blurred image under the same loss; equal gains would indicate the benefit is generic regularization rather than learned de-corruption.","tokens_in":1780,"feed_emoji":"🛡️","tokens_out":1797,"duration_ms":66622,"temperature":0.7,"pith_summary":"The paper tries to establish that overlapping-object perception—detecting or segmenting foreground objects whose contours and textures are partly buried in background clutter—is materially improved by learning from the Fourier magnitude spectrum, not just from spatial features. It claims that the degradation from overlap shows up clearly in the magnitude spectrum, and that a plugin framework called FOAM lets a model re-weight frequency bands to recover texture and contour cues. FOAM's Frequency Spatial Transformer Block (FSTB) combines frequency and spatial self-attention, while its Hierarchical De-Corrupting (HDC) mechanism trains a corrupted-image branch to match the clean branch, teaching the network to suppress background response. If the paper is right, adding FOAM to an existing detector or segmenter is a general way to raise accuracy on security-screening and medical X-ray tasks without changing the task-specific head.","feed_headline":"FOAM boosts DINO by 2.9 AP on overlapping X-ray detection","feed_subtitle":"The same plugin lifts prohibited-item detection, segmentation, and pneumonia detection.","key_machinery":"The Frequency Spatial Transformer Block (FSTB) is the core module: it runs a channel self-attention over spatial features (SDCA), a spatial self-attention over the FFT magnitude spectrum (FDBA) that reweights frequency bands while leaving the phase untouched, and a feed-forward network (FSFN) that fuses the two domains. Cascaded FSTBs form the base branch, which runs at inference. The Hierarchical De-Corrupting (HDC) mechanism is the training-time companion: it applies Gaussian blur, downsampling, or noise to create a corrupted image, runs the same FSTBs on it through shared weights, and uses a KL-based consistent loss to align corrupted features with cleaner base-branch features, with an MSE variant as an alternative. The stated rationale is that suppressing background response increases foreground-to-background contrast at contours, expressed as Theorem I's inequality.","core_discovery":"The central discovery is that the magnitude spectrum of an image is a reliable signal for overlapping-object perception: when contours and textures are degraded by overlap, the magnitude spectrum changes in characteristic ways, while the phase spectrum and spatial appearance stay comparatively stable. FOAM exploits this by decoupling magnitude from phase: FDBA applies self-attention across frequency bands of the magnitude spectrum to reweight texture-carrying bands, the phase is preserved, and an inverse Fourier transform returns a corrected feature map that is fused with spatial features. The HDC mechanism then creates a second branch on a deliberately corrupted image and aligns its features with the clean branch via a KL-divergence consistent loss, so the FSTB learns to suppress background features and sharpen foreground contours. The paper reports that this raises DINO from 64.3 to 67.2 box AP on PIXray-det with ResNet-50, that F-DINO with Swin-L reaches 73.7 AP and 81.7 mAP on OPIXray, and that the same recipe improves instance segmentation and pneumonia detection.","pith_inferences":["Because HDC needs no overlap annotations, FOAM should transfer to any heavily occluded detection domain, such as crowd counting, autonomous driving, or camouflaged objects, provided the magnitude-spectrum degradation assumption holds.","The paper leaves open whether reweighting the magnitude spectrum is the active ingredient or whether any global frequency perturbation would regularize equally well; a controlled swap of phase versus magnitude corruption would settle this.","The HDC mechanism resembles self-supervised denoising, so a natural extension is to use the corruption branch for test-time adaptation on unlabeled degraded images from a target hospital or scanner."],"forward_implications":["Plugging FOAM into a detector or segmenter improves accuracy without changing the task-specific head: F-DINO, F-Mask-R-CNN, and F-Cascade-R-CNN all beat their bases.","FOAM generalizes across CNN and Transformer backbones (ResNet-50, ResNeXt-101, Swin-L) and across detection and instance segmentation architectures including Deformable-DETR, DINO, RT-DETR, Mask2Former, CondInst, Cascade-Mask-R-CNN, and Mask-R-CNN.","Because the corruption branch is used only at training time, inference cost is just the added FSTBs; the paper reports about 1.4M extra parameters and 4.0 GFLOPs for the N=1 DINO configuration.","The consistent loss choice matters: KL-divergence (Type I) gives higher and more stable AP than MSE (Type II), and Gaussian blur is the most effective corruption strategy, with kernel size 3 and sigma 5.","Aligning only higher-level multi-scale layers yields the best results; including low-level layers hurts, since low-level features carry noise."],"supporting_citations":[{"why":"Supplies the PIXray-det and PIXray-seg benchmarks where the main detection and segmentation gains are demonstrated.","marker":"[1]"},{"why":"Provides the OPIXray fine-grained knife benchmark on which F-DINO reaches 81.7 mAP.","marker":"[2]"},{"why":"Is the RSNA pneumonia dataset from which the RLP subset is built for the medical detection comparison.","marker":"[3]"},{"why":"Is the DINO baseline that FOAM improves from 64.3 to 67.2 box AP with ResNet-50.","marker":"[57]"},{"why":"Is the Deformable-DETR architecture underlying DINO and another detector that FOAM improves.","marker":"[45]"},{"why":"Is the fixed high-pass-filter frequency method that FOAM contrasts with and outperforms.","marker":"[12]"},{"why":"Is the SRM-filter frequency-aware method that FOAM contrasts with and outperforms.","marker":"[14]"},{"why":"Is a specialized anti-overlapping DETR detector that F-DINO surpasses on prohibited-item benchmarks.","marker":"[6]"},{"why":"Provides the Fast Fourier Transform used throughout FSTB for magnitude and phase extraction.","marker":"[13]"},{"why":"Supplies the FFT-based frequency-spatial entanglement approach that FOAM extends with magnitude-phase decoupling.","marker":"[32]"}],"fun_headline_variants":["FOAM boosts DINO by 2.9 AP on overlapping X-rays","Frequency domain plugin lifts overlap detection by 2.9 AP","Magnitude spectrum unlocks overlapping object perception","FOAM uses frequency to sharpen overlapping object detection","Frequency-based plugin boosts DINO on X-ray overlap tasks"],"cache_read_input_tokens":29568,"weakest_assumption_plain":"The load-bearing premise is that training a deliberately corrupted branch to match the clean branch suppresses background responses and sharpens foreground contours; if that premise fails, HDC contributes nothing beyond the FSTB.","fun_headline_variants_meta":{"raw":{"variants":["FOAM boosts DINO by 2.9 AP on overlapping X-rays","Frequency domain plugin lifts overlap detection by 2.9 AP","Magnitude spectrum unlocks overlapping object perception","FOAM uses frequency to sharpen overlapping object detection","Frequency-based plugin boosts DINO on X-ray overlap tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000661,"raw_usage":{"total_tokens":3069,"prompt_tokens":1040,"completion_tokens":2029,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":1949}},"tokens_in":656,"tokens_out":2029,"duration_ms":14431,"temperature":1.0,"reasoning_tokens":1949,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:59:43.444980+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure foreground-background contrast in the feature maps of DINO, DINO+FSTB, and FSTB+HDC under identical training budgets. If box AP rises with HDC but average contour contrast in the feature maps does not increase, then the claimed suppression mechanism is not what drives the gain. Alternatively, replace the corrupted-image branch with an unrelated randomly blurred image under the same loss; equal gains would indicate the benefit is generic regularization rather than learned de-corruption.","supporting_citations":[{"cited_title":"Automated segmentation of prohibited items in x-ray baggage images using dense de-overlap attention snake,","cited_arxiv_id":null,"evidence_quote":"Supplies the PIXray-det and PIXray-seg benchmarks where the main detection and segmentation gains are demonstrated."},{"cited_title":"Occluded prohibited items detection: An x-ray security inspection benchmark and de-occlusion attention module,","cited_arxiv_id":null,"evidence_quote":"Provides the OPIXray fine-grained knife benchmark on which F-DINO reaches 81.7 mAP."},{"cited_title":"Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia,","cited_arxiv_id":null,"evidence_quote":"Is the RSNA pneumonia dataset from which the RLP subset is built for the medical detection comparison."},{"cited_title":"Deformable {detr}: Deformable transformers for end-to-end object detection,","cited_arxiv_id":null,"evidence_quote":"Is the Deformable-DETR architecture underlying DINO and another detector that FOAM improves."},{"cited_title":"Feature-aware prohibited items detection for x-ray images,","cited_arxiv_id":null,"evidence_quote":"Is the fixed high-pass-filter frequency method that FOAM contrasts with and outperforms."},{"cited_title":"Fdtnet: Enhanc- ing frequency-aware representation for prohibited object detection from x-ray images via dual-stream transformers,","cited_arxiv_id":null,"evidence_quote":"Is the SRM-filter frequency-aware method that FOAM contrasts with and outperforms."},{"cited_title":"Ao-detr: Anti-overlapping detr for x-ray prohibited items detection,","cited_arxiv_id":null,"evidence_quote":"Is a specialized anti-overlapping DETR detector that F-DINO surpasses on prohibited-item benchmarks."},{"cited_title":"The fast fourier trans- form and its applications,","cited_arxiv_id":null,"evidence_quote":"Provides the Fast Fourier Transform used throughout FSTB for magnitude and phase extraction."},{"cited_title":"Frequency-spatial entanglement learning for camouflaged object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the FFT-based frequency-spatial entanglement approach that FOAM extends with magnitude-phase decoupling."}],"review_version":2}