{"id":"82024d9d-0245-4756-9243-dcad9db2ac57","arxiv_id":"2504.16487","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A domain-adaptive detection framework combining gamma alignment, Poisson fusion augmentation, and noise-robust representation learning reports state-of-the-art results on a new infrared UAV benchmark.","lead":"This paper tests a framework for infrared small target detection that aligns pixel statistics across datasets, blends target patches with diverse backgrounds, and regularizes the model against noise, and it introduces a new UAV benchmark. The authors report state-of-the-art detection accuracy and sharply lower false alarm rates on their benchmark and on IRSTD-1K.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No documented train/test split for RealScene-ISTD; reported gains may come from testing on training images.","rationale":"The reader's weakest assumption — that RealScene-ISTD test statistics and alpha selection may use the same images used for training — is the most load-bearing concern. The central claim is the empirical superiority of the proposed framework in multi-dataset training (Tables II and III). That claim depends entirely on a fair, disjoint evaluation. The paper provides no explicit split, and its language in Section IV is ambiguous ('for Real-Scene training and validate them separately'). Section V-B explicitly trains on RealScene-ISTD and Table II tests on it, while Section V-D selects alpha on the test set. If the same 739 images are used both for training and testing, the reported IoU/Pd/Fa numbers are not evidence of generalization and the comparison to baselines is invalid. Additional internal inconsistencies (Table I vs Table II for SCTransNet; Table VII baseline matching Table II SCTransNet exactly) weaken trust in the numbers as reported. However, the paper provides code and a plausible framework; the flaw is a missing evaluation protocol rather than a demonstrated false result. A concrete check of the released code can settle whether the split exists. Therefore the reader's CONDITIONAL verdict is appropriate and unchanged.","tokens_in":15762,"tokens_out":4888,"duration_ms":44438,"concrete_test":"Inspect the released repository (https://github.com/luy0222/RealScene-ISTD): verify whether the data loader for multi-dataset training includes all 739 RealScene-ISTD images and whether the evaluation script for Table II reads from the same directory. Also check whether alpha=0.6 was chosen using a held-out validation split rather than the test set. If no split exists, rerun the multi-dataset comparison with a fixed 80/20 train/test split (and select alpha on the validation portion); report whether the IoU/Pd/Fa margins over SCTransNet persist.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper never defines a train/test split for RealScene-ISTD. Section IV states the 739 images are used 'for Real-Scene training and validate them separately,' but no split ratio, file list, or protocol is given. Section V-B trains all methods on NUAA-SIRST + IRSTD-1K + RealScene-ISTD, and Table II reports test results on RealScene-ISTD from that same training set. Section V-D selects alpha=0.6 by evaluating on the RealScene-ISTD test set (Table IV). If the 739 images appear in both training and test, the headline improvements (79.32 IoU, 96.83 Pd, 5.40 Fa) are training-set accuracy, not generalization, and the comparison against SCTransNet and others is circular. This concern is reinforced by internal numeric inconsistencies: Table I reports SCTransNet multi-dataset RealScene-ISTD as 77.00 IoU/93.23 Pd/60.51 Fa, while Table II reports 75.01/94.29/53.59; Table VII's baseline row equals Table II's SCTransNet row exactly (75.01/94.29/53.59), suggesting mislabeled or duplicated rows. These inconsistencies do not by themselves refute the method, but they make the reported gains unverifiable without a defined split.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a domain-adaptation framework for infrared small target detection (ISTD) with three main components: Cross-view Channel Alignment (CCA), which gamma-corrects source-domain images to match target-domain gray-level statistics; Cross-view Top-K Poisson Fusion, which pastes target patches into matched background regions; and Noise-guided Representation Learning, which imposes feature consistency between clean and noisy views. The authors also introduce RealScene-ISTD, a 739-image UAV infrared dataset, and report multi-dataset training experiments on NUAA-SIRST, IRSTD-1K, and RealScene-ISTD, claiming superior IoU, Pd, and Fa over seven baselines. The code is released.","tokens_in":15964,"tokens_out":5564,"duration_ms":50323,"significance":"If the reported results held, the paper would make a useful contribution: a new real-scene benchmark for ISTD and a practical combination of gamma alignment, Poisson blending, and noise-robust representation learning that improves cross-dataset generalization. The release of code and the construction of a new annotated dataset are concrete assets. However, the evaluation protocol as written is not verifiable: for RealScene-ISTD there is no documented train/test split, the noise-strength hyperparameter alpha is selected on the final test set, and the target-domain statistics used by CCA may be computed on test images. These issues directly affect the central claim of state-of-the-art generalization, so the paper is not acceptable in its current form.","major_comments":[{"comment":"The manuscript never defines a train/test split for RealScene-ISTD. Section IV states that the 739 images are used 'for Real-Scene training and validate them separately,' but no split ratio, file list, or protocol is given. Section V-B trains all methods on NUAA-SIRST + IRSTD-1K + RealScene-ISTD, and Table II reports test results on RealScene-ISTD from that same training set. If the 739 images appear in both training and testing, the headline improvements (79.32 IoU, 96.83 Pd, 5.40 Fa) are training-set accuracy, not generalization, and the comparison against SCTransNet and other methods is circular. This concern is reinforced by internal numeric inconsistencies: Table I reports SCTransNet multi-dataset RealScene-ISTD as 77.00 IoU/93.23 Pd/60.51 Fa, while Table II reports 75.01/94.29/53.59; Table VII's baseline row equals Table II's SCTransNet row exactly (75.01/94.29/53.59), suggesting mislabeled or duplicated rows. These inconsistencies do not by themselves refute the method, but they make the reported gains unverifiable without a defined split.","section":"Section IV and V-B"},{"comment":"The hyperparameter alpha in LNoise is selected by evaluating on the RealScene-ISTD test set, as shown in Table IV where alpha=0.6 is chosen as optimal, and the same test set is then used to report the final results in Tables II, III, and VII. This is test-set tuning: the improvement attributed to LNoise (e.g., Fa dropping from 47.19 to 29.61 at alpha=0.6, or the full-model Fa of 5.40 in Table VII) is partly a selection artifact. The authors should use a held-out validation split for alpha selection and report final metrics on an untouched test split.","section":"Section V-D, Table IV"},{"comment":"The CCA gamma factor gammaGray is derived from the mean pixel intensity of the target domain XT (Algorithm 1), but the manuscript does not specify whether XT includes the RealScene-ISTD images used for testing. If target-domain statistics are computed over the test images, CCA leaks test-set information into the training data. The paper must specify exactly which images are used to estimate gammaGray and ensure they are disjoint from the test set.","section":"Section III-B and Algorithm 1"},{"comment":"There are multiple internally inconsistent or incorrectly reported numbers. Table I gives SCTransNet multi-dataset RealScene-ISTD as 77.00 IoU / 93.23 Pd / 60.51 Fa, while Table II gives 75.01 / 94.29 / 53.59 for the same setting; Table VII's baseline row is identical to Table II's SCTransNet row. Table V reports a Fa decrease from 47.19 to 29.61 as '↑17.58%', and Table VI reports Fa decreases as '↑4.92%' and '↑3.30%', but a decrease in false alarm rate is an improvement and should be reported with a downward arrow or as a negative percentage change. Table I also contains implausible values for UIU-Net (e.g., Fa 7008.86 with '↓6946.92%'). These inconsistencies make the quantitative claims difficult to audit and must be corrected.","section":"Tables I, V, VI, VII"},{"comment":"All experimental comparisons appear to be single runs with no error bars or significance tests. Several differences are small (e.g., baseline IoU 76.56 vs. alpha=0.4 IoU 76.57 in Table IV, or SCTransNet IoU 76.56 in Table IX vs. 75.01 in Table II), so without multiple seeds it is not possible to determine whether the reported improvements are statistically meaningful. The authors should report mean and standard deviation over at least three runs for all key tables.","section":"Section V-D and all comparison tables"}],"minor_comments":[{"comment":"There are several typos and grammatical errors: 'develope' in the abstract, 'Corss-view' in the Section III-B heading, 'hermal noise' before Eq. (6), 'γGary' in Algorithm 1, and 'Comparations' in Section V-C. These should be fixed throughout.","section":"Abstract and Section III"},{"comment":"Reference [49] appears to be the same paper as reference [24] (both cite the ACM paper by Dai et al.). Please renumber or cite the actual sources correctly.","section":"References"},{"comment":"The ROC curve comparison in Fig. 9 is mentioned but the curves are not described in enough detail; please add axis labels and legend information so that the curves are interpretable.","section":"Section V-C"},{"comment":"Equation (8) omits the averaging/normalization typically used in an MSE loss; please clarify whether the reported LNoise is a mean over all spatial and channel positions or a sum.","section":"Section III-C, Eq. (8)"}],"recommendation":"major_revision","confidential_remarks":"The core issue is the missing train/test split for the newly introduced RealScene-ISTD benchmark. If the authors can release a proper split, keep the test set untouched during alpha selection and CCA statistics estimation, and correct the inconsistent tables, the paper could become viable. I recommend that the editor verify the GitHub repository actually contains the split, the hyperparameter-selection protocol, and the code before any conditional acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers something real: a new real-scene UAV infrared small target dataset (739 annotated images with varied target scales, poses, backgrounds) and a simple, plausible training pipeline for cross-domain ISTD that combines gamma-based channel alignment, SSIM/Top-K Poisson fusion augmentation, and a noise-guided feature consistency loss. The components are all known, but the assembly is coherent, the code is available, and the dataset addresses a genuine gap. If the results hold up, this would be a useful contribution to the ISTD subfield.\n\nThat \"if\" is doing a lot of work, because the experimental protocol as written does not support the headline numbers. The paper never defines a train/test split for RealScene-ISTD. Section IV says the 739 images are used for \"Real-Scene training and validate them separately,\" but there is no split ratio, file list, or protocol. Then Section V-B trains on NUAA-SIRST + IRSTD-1K + RealScene-ISTD and reports test results on RealScene-ISTD. If those are the same images, the 79.32 IoU is a training-set number, not a generalization result. The stress-test is correct: this is a load-bearing flaw.\n\nThe hyperparameter selection compounds the problem. Table IV picks alpha=0.6 by evaluating on the RealScene-ISTD test set. That inflates the LNoise gain. No error bars are provided; every comparison is single-run, which is common in this area but still worth flagging.\n\nThere are also internal numeric inconsistencies that make the paper hard to audit. Table I and Table II give different numbers for SCTransNet under the same multi-dataset condition (77.00 vs 75.01 IoU on RealScene-ISTD). Table VII's baseline row exactly reproduces Table II's SCTransNet row, which suggests a mislabeled or duplicated row. These do not refute the method, but they undercut the claim that the comparisons are apples-to-apples.\n\nThe core idea is not refuted by these problems. The direction of the improvements is consistent, and the proposed strategy makes sense for the stated problem. But the authors need to specify the split, select hyperparameters on a held-out validation set, report multiple runs, and clean up the tables before any conclusion can be drawn.\n\nI would send this to peer review, not desk-reject it: the dataset and the recipe are worth referee time, and the flaws are addressable. I would not cite it as evidence until the protocol is fixed.","headline":"A useful new UAV infrared dataset and a sensible domain-robust training recipe, but the experimental protocol has a circularity problem that makes the headline numbers unverifiable.","tokens_in":16538,"tokens_out":3507,"would_cite":false,"duration_ms":30188,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A domain-adaptive training recipe — channel alignment, Top-K fusion, and noise-guided learning — makes infrared small-target detectors generalize across datasets and achieves the best IoU, Pd, and Fa on a new benchmark.","keywords":["infrared small target detection","domain shift","cross-view channel alignment","Top-K Poisson fusion","noise-guided representation learning","RealScene-ISTD benchmark","multi-dataset training","UAV infrared imagery"],"falsifier":"Release the exact RealScene-ISTD train/test split, choose $\\alpha$ on a held-out validation set, and recompute the multi-dataset comparison; if the IoU, Pd, and Fa margins over SCTransNet shrink to a few points or reverse, the claimed generalization gains are not supported.","tokens_in":15526,"feed_emoji":"🎯","tokens_out":12724,"duration_ms":102511,"temperature":0.7,"pith_summary":"The paper tackles a practical failure mode: infrared small target detectors trained on one collection of images lose accuracy when applied to images from other sensors, weather conditions, or scenes, because the pixel distributions shift between datasets. To counter this, it proposes a training framework with three interlocking parts: Cross-view Channel Alignment (gamma-correcting source images toward the target domain's gray-level distribution), Cross-view Top-K Poisson Fusion (pasting target patches into the most similar background regions to create realistic synthetic samples), and Noise-guided Representation Learning (adding Gaussian noise to inputs and forcing clean and noisy feature maps to match). The paper also introduces RealScene-ISTD, a 739-image benchmark of real infrared drone scenes with diverse targets, backgrounds, poses, and noise. In multi-dataset training, the method reports IoU of 79.32%, a detection probability of 96.83%, and a false alarm rate of 5.40 on RealScene-ISTD; on IRSTD-1K it reports IoU of 72.44%, a detection probability of 95.30%, and a false alarm rate of 14.08, outperforming seven earlier detectors including SCTransNet. If the results hold, one model can be trained on several heterogeneous infrared datasets at once instead of being retuned per domain.","feed_headline":"New recipe lifts infrared small-target detection to 79.32 IoU","feed_subtitle":"Channel alignment, Top-K fusion, and noise-guided training cut false alarms to 5.40 on a new real-scene UAV benchmark.","key_machinery":"The load-bearing machinery is the joint training objective $L_{\\text{total}} = L_{\\text{BCE}} + L_{\\text{Noise}}$ applied on top of the Cross-view Top-K Fusion samples. $L_{\\text{BCE}}$ is the standard binary cross-entropy segmentation loss; $L_{\\text{Noise}} = \\|F^P_{\\text{Global}} - F^N_{\\text{Global}}\\|^2$ is the MSE between the global multi-scale features of clean and noise-added versions of the same image, computed from a shared-weight encoder-decoder with Attention Features Fusion. The other two pieces feed this objective: Cross-view Channel Alignment computes a gamma factor $\\gamma_{\\text{Gray}}$ from the mean gray-levels of source and target domains so that $O_{\\text{Gray}}=(I_{\\text{Gray}}/255)^{1/\\gamma_{\\text{Gray}}}$, while Cross-view Top-K Poisson Fusion uses SSIM to find the K most similar background regions for each target patch and blends them with Poisson gradient-domain fusion. Together they create a training distribution that is aligned across domains, richer in target-background combinations, and robust to the thermal noise that varies between sensors.","core_discovery":"The central claim is that domain shift, not model capacity, is the main obstacle to generalizable infrared small target detection, and that the obstacle can be lowered by a specific training recipe. The paper shows that simply merging multiple datasets degrades existing detectors — for example, UIU-Net's false alarm rate explodes to 7008.86 — and that its three proposed mechanisms reverse this degradation. Cross-view Channel Alignment maps the source domain's gray-level distribution toward the target domain through a per-channel gamma factor, so multi-dataset training sees more consistent statistics. Cross-view Top-K Poisson Fusion crops real target patches and blends them into the K most SSIM-similar background regions, producing synthetic samples that preserve realistic edges and expand the target-background combinations the model sees. Noise-guided Representation Learning adds zero-mean Gaussian noise with variance $\\alpha^2$ and minimizes the MSE between the global feature maps of clean and noisy copies, forcing the encoder to retain noise-invariant target structure. With the best $\\alpha = 0.6$, the framework reports 79.32 IoU, 96.83 Pd, and 5.40 Fa on RealScene-ISTD and 72.44 IoU, 95.30 Pd, and 14.08 Fa on IRSTD-1K, beating SCTransNet and six other methods.","pith_inferences":["The same gamma-based channel alignment idea could transfer to other small-object or low-contrast detection tasks, such as medical imaging or remote sensing, because it makes no assumption about target type.","The Top-K fusion step's fixed SSIM top-K selection could be made adaptive, with the number of fused regions or the similarity metric learned per domain; the paper does not explore this variant.","The noise-level sweep suggests a performance plateau around $\\alpha = 0.6$; a natural extension would be to anneal $\\alpha$ during training rather than keep it constant.","The paper's framing that multi-dataset training degrades prior methods rests on one particular dataset combination; re-running the same comparison on other combinations would show how general the degradation really is."],"forward_implications":["Multi-dataset training becomes a workable strategy for infrared small target detection: merging heterogeneous datasets no longer collapses accuracy, so models can be trained once on several sources rather than tuned per sensor.","The alignment and fusion modules are portable: adding CCA and Top-K Fusion to a different detector (ISTDU-Net) raises its IoU from 74.40 to 76.90 and cuts its false alarm rate from 75.35 to 39.69.","The noise-guided consistency loss is the largest single lever on false alarms: on RealScene-ISTD it drops the false alarm rate from 47.19 to 29.61 while raising IoU from 76.56 to 77.69.","The released RealScene-ISTD benchmark gives the community a common testbed for cross-domain generalization, with the same multi-dataset training protocol used for all comparisons.","The full framework delivers the paper's headline numbers: 79.32 IoU, 96.83 Pd, and 5.40 Fa on RealScene-ISTD, and 72.44 IoU, 95.30 Pd, and 14.08 Fa on IRSTD-1K, ahead of SCTransNet and six other state-of-the-art detectors."],"supporting_citations":[{"why":"The main comparison baseline; its multi-dataset IoU, Pd, and Fa numbers are the ones the proposed method claims to beat.","marker":"[27]"},{"why":"Supplies the NUAA-SIRST dataset used alongside IRSTD-1K and RealScene-ISTD in the joint training protocol.","marker":"[49]"},{"why":"Supplies the IRSTD-1K dataset used for joint training and as the independent test benchmark in Table III.","marker":"[50]"},{"why":"ACM is one of the seven re-trained comparison methods in Tables II and III.","marker":"[24]"},{"why":"DNA-Net is a comparison baseline whose attention-fusion architecture motivates the multi-scale feature design.","marker":"[21]"},{"why":"UIU-Net provides the clearest failure case motivating the work, with its false alarm rate exploding to 7008.86 under multi-dataset training.","marker":"[26]"},{"why":"ISTDU-Net is the alternative detector used in Table VIII to show CCA and Top-K Fusion transfer across backbones.","marker":"[54]"},{"why":"ALCNet is a comparison baseline for detection metrics and efficiency in Table IX.","marker":"[20]"},{"why":"RDIAN is a comparison baseline re-trained in the multi-dataset protocol.","marker":"[25]"},{"why":"ResNet is the backbone of the U-Net-based detector, so the main experiments depend on its feature extraction.","marker":"[51]"}],"fun_headline_variants":["Cross-view learning tames domain shift in infrared small-target detection","Domain shift defeats IR target detectors; cross-view recipe fixes it","Real-scene IR benchmark with cross-view alignment reaches 79.32 IoU","Three cross-view mechanisms cut IR target false alarms from 7008 to 5.4","New real-scene ISTD dataset and cross-view learning boost IoU to 79.32"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume that the RealScene-ISTD images used to estimate target-domain statistics and to pick the noise strength $\\alpha = 0.6$ are disjoint from the images used to report the final test metrics, so that a genuine train/test split exists for the new benchmark.","fun_headline_variants_meta":{"raw":{"variants":["Cross-view learning tames domain shift in infrared small-target detection","Domain shift defeats IR target detectors; cross-view recipe fixes it","Real-scene IR benchmark with cross-view alignment reaches 79.32 IoU","Three cross-view mechanisms cut IR target false alarms from 7008 to 5.4","New real-scene ISTD dataset and cross-view learning boost IoU to 79.32"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000739,"raw_usage":{"total_tokens":3354,"prompt_tokens":1051,"completion_tokens":2303,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":2200}},"tokens_in":667,"tokens_out":2303,"duration_ms":15914,"temperature":1.0,"reasoning_tokens":2200,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:01:53.556734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Release the exact RealScene-ISTD train/test split, choose $\\alpha$ on a held-out validation set, and recompute the multi-dataset comparison; if the IoU, Pd, and Fa margins over SCTransNet shrink to a few points or reverse, the claimed generalization gains are not supported.","supporting_citations":[{"cited_title":"Sctransnet: Spatial- channel cross transformer network for infrared small target detection,","cited_arxiv_id":null,"evidence_quote":"The main comparison baseline; its multi-dataset IoU, Pd, and Fa numbers are the ones the proposed method claims to beat."},{"cited_title":"Asymmetric contextual modulation for infrared small target detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the NUAA-SIRST dataset used alongside IRSTD-1K and RealScene-ISTD in the joint training protocol."},{"cited_title":"Isnet: Shape matters for infrared small target detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the IRSTD-1K dataset used for joint training and as the independent test benchmark in Table III."},{"cited_title":"Istdu-net: Infrared small-target detection u-net,","cited_arxiv_id":null,"evidence_quote":"ISTDU-Net is the alternative detector used in Table VIII to show CCA and Top-K Fusion transfer across backbones."}],"review_version":1}