{"id":"97331ce2-466c-4f84-88a2-6ba0a0d9592e","arxiv_id":"2504.20602","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"PLUSNet combines Fourier-based feature purification, multi-criteria label assignment, and a frequency-decoupled head to improve small object detection by about 3 AP on SODA-D.","lead":"Small object detection gets a new three-part pipeline, PLUSNet, that purifies low-level features with Fourier filtering, assigns training samples using three criteria instead of IoU alone, and gives classification and regression separate frequency-tuned heads. The full system gains about 3 AP on the SODA-D driving benchmark and a larger gain on AI-TOD aerial images, at the cost of doubled computation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set hyperparameter selection may inflate the reported AP gains; a held-out validation re-run is needed before the headline improvements can be trusted.","rationale":"The reader's weakest_assumption correctly identifies the HFP premise as a soft spot: Table III shows HFP alone decreases APeS from 13.8 to 12.6, so the purification benefit for the smallest objects only appears after MCLA and FDHead are added. That is a genuine concern about the internal narrative, and it deserves the conditional verdict. However, a more load-bearing issue sits one level earlier: the empirical evidence for the central claim is drawn entirely from test-set comparisons with hyperparameters chosen on that same test set. Since the paper reports no validation split, no multiple seeds, and no error bars, the headline improvements could be artifacts of selection and stochastic variation rather than genuine gains. This does not change the verdict: the paper should still be accepted conditionally, with the condition being a validation-based replication or at minimum a clarification of the evaluation protocol and release of code. The concrete test above would settle whether the concern lands; if the validation-tuned AP holds, the central claim survives. My agreement with the reader is partial because I share their HFP worry but believe the test-set tuning is the most load-bearing vulnerability, and it is distinct from the one they emphasized.","tokens_in":18707,"tokens_out":4809,"duration_ms":49033,"concrete_test":"Run PLUSNet and the Faster R-CNN baseline on SODA-D and AI-TOD with a proper train/validation/test split: tune all HFP/MCLA/FDHead hyperparameters (µ, ω, r, λ2, λ3, Dl, Dh) only on the validation split, freeze them, then evaluate on the test split, repeating with at least 3 random seeds and reporting mean±std. If the resulting SODA-D AP is not at least ~1 AP above the baseline and comparable to CFINet's 30.7, the reported gains are at least partly an artifact of test-set tuning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PLUSNet consistently achieves large improvements (Table I: 32.0 AP vs 28.9 baseline and 30.7 CFINet on SODA-D; Table II: 21.7 vs 11.1 on AI-TOD). The load-bearing condition is that these numbers reflect genuine generalization. The paper does not use a validation split: every hyperparameter (µ, ω, r in Tables IV/V, λ2/λ3 in Table VI, Dl/Dh in Table VII, plus the choice of which frequency components in Table VIII) is selected by comparing AP on the SODA-D test set, and the same test set is then used for the final comparison. With roughly ten free hyperparameters and multiple ablation tables, this is a textbook case of selection on the test set, which can inflate performance and make the comparison against baselines unfair, since the baselines are not given the same test-set tuning. Moreover, all results are single runs with no seeds or error bars, so the 1.3 AP advantage over CFINet and the 3.1 AP over baseline could be within run-to-run variance. The HFP-specific concern raised by the reader is real (HFP alone lowers APeS from 13.8 to 12.6 in Table III), but it is secondary: even if HFP works as claimed, the reported numbers would be unreliable due to test-set selection. If the validation-tuned AP drops materially, the central claim fails.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PLUSNet, a small-object detection framework composed of three modules: a Hierarchical Feature Purifier (HFP) that high-pass filters low-level FPN features in the Fourier domain, a Multiple Criteria Label Assignment (MCLA) that augments IoU with position-offset and shape-constraint criteria, and a Frequency Decoupled Head (FDHead) that feeds low-frequency features to classification and high-frequency features to regression. Experiments on SODA-D, AI-TOD, and a COCO subset reportedly show consistent gains over Faster R-CNN and prior state-of-the-art methods, with the largest claimed improvements being 32.0 AP vs. 28.9 AP baseline on SODA-D and 21.7 AP vs. 11.1 AP baseline on AI-TOD. The paper also includes ablations for each module and a simulation experiment showing that MCLA assigns more positive samples to small objects than MaxIoU strategies.","tokens_in":18952,"tokens_out":2962,"duration_ms":30153,"significance":"If the reported results hold, the paper would make a useful empirical contribution: each module is simple, the overall design is clearly motivated by the detection pipeline, and the modules are claimed to be plug-and-play across detectors. The simulation experiment in Section IV-D is a nice sanity check that directly demonstrates the label-imbalance problem MCLA addresses. However, the significance is tempered by the fact that the central empirical claim rests on test-set tuning and single-run comparisons, and the computational overhead of the Fourier transforms is substantial (481.0 GFLOPs vs. 206.7 for the baseline in Table X). The paper is reproducible in principle because the method is described with enough detail to re-implement, but no code is released and no variance estimates are provided.","major_comments":[{"comment":"All hyperparameters are selected by comparing AP on the SODA-D test set, and the final comparison in Table I is also on that same test set. Specifically, µ and ω (Table IV), r (Table V), λ2 and λ3 (Table VI), Dl and Dh (Table VII), and the frequency-component choice (Table VIII) are all tuned on the test set. With roughly ten free choices, this constitutes selection on the test set, which can inflate the reported 32.0 AP and make the comparison against CFINet (30.7) and the baseline (28.9) potentially unfair, since the competing methods are not given the same test-set tuning. The authors should instead fix all hyperparameters using a held-out validation split (or cross-validation), report the validation choices, and then evaluate on the test set exactly once. Without this, the headline improvements cannot be trusted as unbiased estimates of generalization.","section":"Section IV-C, Tables IV–VIII"},{"comment":"All reported numbers are single runs with no seeds, error bars, or significance tests. The advantage over CFINet on SODA-D is only 1.3 AP (32.0 vs. 30.7), which may be within run-to-run variance for a 12-epoch schedule on a dataset of this size. The authors should report mean and standard deviation over at least three independent runs for the baseline, the prior state of the art, and PLUSNet, and state whether the differences are statistically meaningful. This is necessary to support the claim of 'significant and consistent improvements' in the abstract.","section":"Tables I–III"},{"comment":"The ablation contradicts the stated role of HFP. Adding HFP alone raises overall AP from 28.9 to 30.0 but lowers APeS (the 0–12 pixel objects) from 13.8 to 12.6. The text in Section IV-C acknowledges this ('the detection accuracy for extremely small objects is adversely affected'), which means the paper's claimed mechanism—that HFP purifies low-level features to benefit small objects—is not directly supported by the data. The final APeS of 15.4 is only achieved after MCLA and FDHead are added. The authors should either present evidence that the HFP-induced drop is due to insufficient sample count rather than feature damage, or reframe HFP as a global feature rebalancer whose benefit for small objects is conditional on the downstream modules.","section":"Section IV-C, Table III"},{"comment":"The computational cost is a major weakness that is underplayed. Table X shows PLUSNet at 481.0 GFLOPs versus 206.7 for the Faster R-CNN baseline and 227.1 for CFINet, i.e., roughly 2.3× the baseline and 2.1× the state of the art on SODA-D. While the paper includes a 'Limitation' paragraph, the abstract and conclusion still present PLUSNet as a straightforward improvement without qualifying the cost. At minimum, the authors should report inference speed (e.g., FPS) and explicitly compare the accuracy-per-FLOP trade-off against CFINet, and should temper the 'plug-and-play' claim to note the computational overhead.","section":"Section IV-F, Table X"}],"minor_comments":[{"comment":"In the text after Algorithm 1, 'SIoU , SPOC , and SPOC' should read 'SIoU , SPOC , and SSCC'.","section":"Section III-C"},{"comment":"The phrase '⊙ denotes the element-wise production' should be 'element-wise product'.","section":"Section III-B, Eq. (3)"},{"comment":"The heading 'Comparision with State-of-the-art Methods' contains a typo; it should be 'Comparison'.","section":"Section IV-B heading"},{"comment":"The table title says 'COCO mini-val SET'. It is unclear what 'mini-val' denotes; please clarify whether this is a random subset of the COCO val set, and describe its size, to allow reproducibility.","section":"Section IV-D, Table IX"},{"comment":"The description of SODA-D masking in the Fig. 5 caption is confusing: 'the SODA-D dataset applies masking to confine the detection objects within a sufficiently small area. This operation directly removes the pixels of larger objects.' Please clarify whether this is part of the dataset's evaluation protocol or a preprocessing step, and whether it affects the quantitative results.","section":"Section IV-E"},{"comment":"The phrase 'The inspiring work [44]' is subjective; please rephrase as 'The work [44]' or 'The influential work [44]'.","section":"Section II-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like a strong conference paper that has been reformatted as a journal submission. The central idea is coherent and the ablations are internally consistent, but the test-set tuning issue is exactly the kind of methodological concern that a careful journal reviewer should enforce. If the authors re-run with validation-based selection and multi-seed reporting, the paper could be a solid candidate for acceptance; as is, the headline numbers are not fully supported. I would not reject, because the concern is addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the method is a sensible three-module pipeline — HFP for frequency-domain feature purification, MCLA for multi-criteria label assignment, FDHead for task-decoupled classification/regression — and the combination is genuinely new even though each ingredient has clear antecedents. Second, the reported gains (32.0 vs 28.9 AP on SODA-D, 21.7 vs 11.1 on AI-TOD) are not yet trustworthy, because essentially every hyperparameter is tuned directly on the SODA-D test set and the same test set is then used for the final comparison. That is selection on the test set, and the baselines are not given the same tuning. No seeds or error bars are reported either, so we cannot tell whether the 1.3 AP edge over CFINet is signal or noise.\n\nWhat the paper does well: the writing is clear, the ablations are systematic, and the authors honestly disclose the doubled compute in their limitation section. The simulation experiment for label assignment is a nice touch, and the method is described in enough detail to re-implement from the equations alone. I also want to correct one thing in the reader's take: the paper does cite Double-Head R-CNN (reference [19]) in the FDHead discussion, so the 'omitted' concern is not accurate.\n\nThe soft spots are real but not fatal. The test-set tuning is the load-bearing issue; a held-out validation split or a re-run with frozen hyperparameters would settle it. The HFP module alone actually hurts APeS in the ablation (13.8 to 12.6), so the claimed purification benefit for the smallest objects only appears after MCLA and FDHead are added. That is fine as a synergy claim, but it means the mechanism is less clean than the narrative suggests. The COCO generalization experiment uses a mini-val set, which is weaker than full COCO val. And of course no code is released, so the numbers cannot be independently checked.\n\nWho is this for? Anyone working on small object detection, especially in autonomous driving or aerial imagery. The paper deserves a serious referee: the idea is coherent, the evaluation is structured, and the gains are large enough to warrant independent replication. But I would ask the authors for code, error bars, and a validation-based hyperparameter selection before trusting the headline numbers. If those come through, this could be a useful contribution; if not, the reported improvements are likely inflated.\n\nMy recommendation: send it to peer review with a request for major revision — specifically, re-run the ablations on a validation split and report variance. It is not a desk reject, but it is not ready as-is.","headline":"A plausible, clearly-written small-object detection pipeline whose headline gains are real only if the test-set hyperparameter tuning and missing error bars do not hide them; worth sending to peer review but the numbers should be treated as provisional.","tokens_in":19553,"tokens_out":1622,"would_cite":false,"duration_ms":19232,"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":"PLUSNet's three modules lift small-object AP to 32.0 on SODA-D and 21.7 on AI-TOD.","keywords":["small object detection","feature pyramid network","frequency domain","label assignment","detection head","Fourier transform","SODA-D","AI-TOD"],"falsifier":"The quickest direct test is to run only HFP, the highpass purification module, on AI-TOD with the settings tuned on SODA-D; the paper's own ablation predicts the tiny-object split will worsen, so an improvement would falsify the purification premise. A second check is to grid-search the filtering strength on AI-TOD and see whether the SODA-D-optimal value still transfers.","tokens_in":18458,"feed_emoji":"🔍","tokens_out":13785,"duration_ms":120430,"temperature":0.7,"pith_summary":"This paper claims that small-object detection is held back by three distinct weaknesses along the detection pipeline, and that fixing all three together yields gains none of the modules achieves alone. It proposes PLUSNet, which purifies low-level features with frequency-domain highpass filtering, replaces IoU-only label assignment with a three-criteria score, and splits the head so classification reads low-frequency features while regression reads high-frequency ones. On the SODA-D driving benchmark, PLUSNet raises average precision (AP) from 28.9 for the Faster R-CNN baseline to 32.0, and on the AI-TOD aerial benchmark from 11.1 to 21.7. The same modules also improve COCO multi-scale detection, supporting the paper's claim that the approach generalizes beyond tiny-object benchmarks.","feed_headline":"Three-part fix nearly doubles tiny-object detection accuracy","feed_subtitle":"Purification, multi-criteria labeling, and a decoupled head push Faster R-CNN from 11.1 to 21.7 AP on AI-TOD.","key_machinery":"The load-bearing mechanism is the trio of modules PLUSNet inserts at three pipeline stages. HFP filters low levels of the feature pyramid in the Fourier domain, with mask width set by $\\mu (r-l)/r$ and the purified map formed as $\\tilde{X}_l = \\mathrm{IFFT}(M \\odot \\mathrm{FFT}(X_l)) \\cdot \\omega + X_l$; this residual highpass filter is what 'purifies' features. MCLA scores each proposal as a normalized weighted sum of IoU, a position offset score $S_{\\mathrm{POC}} = (1+\\sqrt{20 E_1^{\\mathrm{norm}}})^{-1}$, and a shape constraint score $S_{\\mathrm{SCC}} = (1+\\sqrt{0.25 E_2})^{-1}$, which supplies more positive samples to tiny objects. FDHead splits the region-of-interest feature into low-frequency and high-frequency components, feeding the low-frequency part through fully connected layers for classification and the high-frequency part through convolution blocks with global average pooling for regression. The paper argues these three mechanisms are complementary, and that their combination, not any single module, produces the reported accuracy.","core_discovery":"The central claim is that the standard two-stage detection pipeline damages small objects at every stage, so optimizing the whole pipeline works better than optimizing any isolated stage. In the upstream, feature pyramid top-down fusion injects low-frequency semantic noise into the low-level features that small objects need; the Hierarchical Feature Purifier applies an FFT-based highpass mask with a residual to those levels. In the midstream, IoU-only assignment under-supplies tiny objects with positive training samples because the ratio is size-sensitive; the Multiple Criteria Label Assignment adds a position-offset criterion and a shape-constraint criterion to the IoU score. In the downstream, a shared head forces classification and regression to use the same representation; the Frequency Decoupled Head gives classification fully connected layers on low-frequency features and regression convolutions on high-frequency features. The paper's evidence is the full model's gains on SODA-D, AI-TOD, and COCO, plus ablations showing each module contributes and that the final small-object benefit requires their combination.","pith_inferences":["If the frequency-band premise transfers, the HFP mask should be learnable per level and per dataset rather than fixed by $\\mu$ and $r$; a testable prediction is that the optimal cutoff tracks the dataset's object-size distribution, with AI-TOD's roughly 12.8-pixel objects needing a different pass band than SODA-D's roughly 20.3-pixel objects.","The MCLA simulation counts how many positive samples each assigner produces, but not whether those extra samples actually improve localization; measuring regression error on the newly assigned tiny-object samples would show whether the gain comes from more samples or from more balanced training.","Because the paper reports a large computational cost increase (206.7 to 481.0 GFLOPs) and labels it a limitation, a practical follow-up would replace the explicit FFT/IFFT loops with a learned frequency filter or a distilled model that skips the transform at inference; the paper itself suggests learning to work directly on the transformed features."],"forward_implications":["On SODA-D, PLUSNet with a ResNet-50 backbone reaches 32.0 AP versus 28.9 for the Faster R-CNN baseline and 30.7 for the previous best method CFINet.","On AI-TOD, PLUSNet reaches 21.7 AP versus 11.1 for the baseline, and adding the modules to DetectoRS pushes AP to 25.2.","On the COCO mini-val set, adding the modules to Faster R-CNN, Mask R-CNN, Libra R-CNN, and Guided Anchoring improves overall mAP and the small-object metric, so the claimed benefit is not limited to purpose-built tiny-object datasets.","In the SODA-D ablation, HFP alone lowers the metric for 0-to-12-pixel objects (APeS) from 13.8 to 12.6; the full gain appears only after MCLA and FDHead are added, eventually reaching 15.4 APeS."],"supporting_citations":[{"why":"supplies the fast Fourier transform that HFP and FDHead use to split features into frequency bands.","marker":"[11]"},{"why":"defines the feature pyramid network whose low levels HFP purifies.","marker":"[12]"},{"why":"is the Faster R-CNN baseline on which PLUSNet is built and compared.","marker":"[13]"},{"why":"provides the SODA-D benchmark and its four size-based AP splits used for the main comparison.","marker":"[20]"},{"why":"provides the AI-TOD aerial benchmark used for the second main comparison.","marker":"[21]"},{"why":"is the MS COCO dataset used to test whether the modules generalize to multi-scale detection.","marker":"[22]"},{"why":"introduces the Gaussian receptive-field label-assignment baseline that MCLA is contrasted with and outperforms.","marker":"[38]"},{"why":"is the prior coarse-to-fine small-object detector whose SODA-D result PLUSNet surpasses.","marker":"[40]"},{"why":"supplies the COCO baseline results used in the plug-and-play generalization table.","marker":"[58]"}],"fun_headline_variants":["Three-part pipeline fix nearly doubles small-object AP","Holistic pipeline design boosts tiny-object detection by 2x","Purify, label, utilize: whole-pipeline fix lifts small-object AP","Optimizing full detection pipeline nearly doubles tiny-object accuracy","PLUSNet: end-to-end pipeline fix improves small-object detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method depends on the premise that the lowest-level feature maps in the feature pyramid contain harmful low-frequency semantic noise, and that removing it with highpass filtering plus a residual preserves the small-object detail; the ablation table shows this module alone actually hurts the smallest objects, with the 0-to-12-pixel metric APeS falling from 13.8 to 12.6, so the purification benefit appears only after the other two modules are added.","fun_headline_variants_meta":{"raw":{"variants":["Three-part pipeline fix nearly doubles small-object AP","Holistic pipeline design boosts tiny-object detection by 2x","Purify, label, utilize: whole-pipeline fix lifts small-object AP","Optimizing full detection pipeline nearly doubles tiny-object accuracy","PLUSNet: end-to-end pipeline fix improves small-object detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":3037,"prompt_tokens":988,"completion_tokens":2049,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":1964}},"tokens_in":604,"tokens_out":2049,"duration_ms":12738,"temperature":1.0,"reasoning_tokens":1964,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:24:19.663642+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The quickest direct test is to run only HFP, the highpass purification module, on AI-TOD with the settings tuned on SODA-D; the paper's own ablation predicts the tiny-object split will worsen, so an improvement would falsify the purification premise. A second check is to grid-search the filtering strength on AI-TOD and see whether the SODA-D-optimal value still transfers.","supporting_citations":[{"cited_title":"Feature pyramid networks for object detection,","cited_arxiv_id":null,"evidence_quote":"defines the feature pyramid network whose low levels HFP purifies."},{"cited_title":"Tiny object detection in aerial images,","cited_arxiv_id":null,"evidence_quote":"provides the AI-TOD aerial benchmark used for the second main comparison."},{"cited_title":"Small object detection via coarse-to-fine proposal generation and imitation learning,","cited_arxiv_id":null,"evidence_quote":"is the prior coarse-to-fine small-object detector whose SODA-D result PLUSNet surpasses."}],"review_version":1}