{"id":"10fadffb-e573-4ddd-8643-013b65a37c3c","arxiv_id":"2505.02441","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"MSFNet-CPD fuses super-resolved images with per-species text and beats YOLOv9 in mAP, but the text descriptions are class labels in disguise.","lead":"This paper pairs pest photos with text descriptions written from the species labels and adds super-resolution, reporting higher detection accuracy than unimodal models on new versions of the IP102 dataset. The result is not yet convincing because the text leaks the answer, and the promised code and data have not been released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time text input is derived from ground-truth species labels, so the reported cross-modal gains may be a label-leak artifact rather than evidence of fusion.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: per-species text descriptions at test time encode the ground-truth class, making the multimodal comparison circular. My independent reading of Sections II-F, III-C, Table II, Figure 3, and Table VI confirms this is not a minor protocol nuance but the main source of the reported improvement. The 10.42 mAP drop when text is removed in Table VI is consistent with label leakage rather than with a learned semantic fusion benefit. Because the paper's headline result is the SOTA comparison in Section III-C and the abstract, and because that comparison is invalidated by the class-conditional text input, the REJECT verdict stands. I do not see a need to move the verdict; the concern fully supports the existing rejection. I did not find an independent load-bearing issue beyond the label leak, although the missing code and datasets and uncontrolled comparisons would further hinder verification if the leak were fixed.","tokens_in":11051,"tokens_out":2705,"duration_ms":34188,"concrete_test":"Re-run the CTIP102 test evaluation with the text input decoupled from the ground-truth label. Specifically, evaluate MSFNet-CPD under three conditions: (1) true per-image text as in Table V; (2) text from a randomly assigned different class for each image; (3) no text input. If condition (2) or (3) drops mAP to roughly the 35.64% level of the w/o-text ablation, while condition (1) stays near 46.06%, the test-time text is acting as a label leak and the central claim fails. Additionally, train the same MSFNet-CPD backbone on unimodal IP102 with identical hyperparameters and compare to Table V to verify the fair baseline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that cross-modal fusion with text descriptions yields state-of-the-art pest detection rests on an evaluation protocol in Sections II-F and III-C where each test image is paired with the text description of its true species. Figure 3 gives, per species, a description such as 'The adult body of the rice leaf roller...', and Table II shows per-image text counts for STIP102 and CTIP102. At inference, the model therefore receives the answer class through the text modality. Comparing MSFNet-CPD on CTIP102 (46.06 mAP) against YOLOv9 on unimodal IP102 (42.32 mAP) is not a controlled test of multimodal fusion: the text contains the label, and the unimodal baseline cannot see it. The ablation in Table VI corroborates the leak: removing text drops mAP by 10.42 points, roughly the gap claimed over YOLOv9. This does not show that jointly modeling image and text is beneficial for pest detection; it shows the model can exploit a class-identifying side channel. The claimed cross-modal advantage and the 'consistently outperforms' conclusion in the abstract are therefore not supported by the reported experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MSFNet-CPD, a multi-scale cross-modal fusion network for crop pest detection. The method combines super-resolved and original images with per-species text descriptions, and the authors introduce four datasets (HIP102, STIP102, CTIP102, MTIP102) derived from IP102, plus an ACIE data-enhancement algorithm for multi-target images. The central claim is that MSFNet-CPD consistently outperforms state-of-the-art unimodal detectors on these benchmarks, with the best result reported on the CTIP102 dataset (mAP 46.06, mAP50 92.18, mAP75 40.18).","tokens_in":11251,"tokens_out":3167,"duration_ms":39085,"significance":"If the evaluation protocol were sound, the paper would make a useful contribution by introducing multimodal pest-detection benchmarks and demonstrating a cross-modal benefit. The datasets, particularly STIP102 and CTIP102, could be of interest to the community, and the ACIE algorithm addresses a practical multi-target scenario. However, the central claim of cross-modal superiority rests on an evaluation setup in which the test-time text input is generated from the ground-truth species label. This confound, together with uncontrolled comparisons against published baselines, means the reported gains are not evidence for the paper's central conclusion. The strength of the contribution therefore depends entirely on whether the protocol can be fixed, which the current manuscript does not demonstrate.","major_comments":[{"comment":"The comparison in Table IV versus Table V is not controlled. Table IV reports YOLOv9 on unimodal IP102 at 42.32 mAP, while Table V reports MSFNet-CPD on CTIP102 at 46.06 mAP. These differ in the dataset (IP102 vs CTIP102), in the training data (the multimodal model also sees text, which leaks the label), and possibly in the train/validation/test split: Section III-A states only that datasets are divided 'according to 8:1:1' with no random seed or confirmation that the split matches that used for the published IP102 baselines in Table IV. No experiment runs MSFNet-CPD on unmodified IP102 under the same training conditions as the baselines, and no error bars or repeated runs are reported. The 0.32-point mAP75 difference (Table V 40.18 vs Table IV 39.86) is well within run-to-run noise for object detectors, so the claim of 'significantly outperforms this state-of-the-art model across all metrics' is not supported by the presented evidence.","section":"§II-F, Figure 3; §III-C; Table VI"},{"comment":"There is an internal inconsistency in the reported mAP75 values for the full model on CTIP102. Table V lists the CTIP102 row as P=82.15, F1=78.21, mAP=46.06, mAP50=92.18, mAP75=40.18, while Table VI lists 'Our model' with mAP=46.06, mAP50=92.18, mAP75=39.86. The same setup should produce the same number; if the difference is due to a different dataset split or random seed, that should be stated. As written, this discrepancy undermines the reliability of the numerical comparisons, especially since the mAP75 value in Table VI coincides exactly with the YOLOv9 mAP75 in Table IV.","section":"Table V vs Table VI"},{"comment":"The ACIE-generated MTIP102 results are not compared against any baseline detector on the same data. Table V reports only MSFNet-CPD on MTIP102 (mAP 22.33), and Section III-E attributes the drop to task complexity. Without a unimodal baseline trained on MTIP102, the reader cannot tell whether the low score is due to the difficulty of the multi-target task or to a deficiency of the proposed method on that data.","section":"§III-A and Table II"}],"minor_comments":[{"comment":"Section II-G states parameter values for ACIE (B=580, T=820, R=4, num=10000) without any justification or sensitivity analysis; the choice of these values may materially affect the MTIP102 results.","section":"§III-C"}],"recommendation":"reject","confidential_remarks":"The central evaluation is invalid because the test-time text input is generated from ground-truth labels, and the comparison against unimodal baselines is not controlled. This is not a matter of presentation; it affects the main claim of the paper. The label-leak confound is confirmed by the ablation study, which shows a 10-point drop when text is removed. I recommend rejection. If the authors want to pursue this line, they would need to redesign the evaluation protocol so that text does not carry the answer (e.g., class-conditioned prompting evaluated on held-out classes) and re-run all comparisons on identical data and splits, with error bars. The dataset construction itself may still be of interest as a resource, but the current paper does not establish that."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's central claim—that fusing per-species text descriptions with images improves pest detection beyond unimodal baselines—doesn't survive its own evaluation protocol. The text inputs are built from the same class labels that the model is supposed to predict (Section II-F, Figure 3). At test time each image is paired with a description of its true species, so the text channel leaks the answer. The 10.42 mAP drop when text is removed (Table VI) is consistent with a label-leak side channel, not evidence for cross-modal reasoning. Comparing CTIP102 (46.06 mAP) to YOLOv9 on unimodal IP102 (42.32) isn't controlled: one model sees the class in the text, the other doesn't. The abstract's 'consistently outperforms' is not backed by these experiments.\n\nWhat's actually new: the paper contributes four datasets derived from IP102 (HIP102, STIP102, CTIP102, MTIP102) with per-class text descriptions, plus a composed architecture—ESRGAN super-resolution, BERT text encoding, Transformer-based cross-modal fusion, and a YOLOv4 head. The ACIE augmentation for generating multi-target images is a reasonable idea, and the paper includes ablations and Grad-CAM visuals. The writing is readable.\n\nSoft spots, in proportion: the label leak is load-bearing and fatal to the main claim. Second, the comparison to YOLOv9 has no error bars or significance tests, and the data split isn't clearly matched to the published baselines. Third, there's an internal inconsistency—Table V lists CTIP102 mAP75 as 40.18 while Table VI gives the same setup 39.86. Fourth, minor but telling: the ACIE pseudocode loops to the annotation step on overlap instead of reselecting a point, and reference [29] doesn't match RefineDet. Finally, datasets and code are promised but not released.\n\nThis paper is a useful cautionary example for a reading group—it shows how easy it is to build a multimodal benchmark and then accidentally leak the label through a text channel. But as a research contribution, the quantitative results are unreliable. My recommendation: desk reject, or if the authors redo the evaluation with text that doesn't reveal the class identity and release the data, a revised version might deserve a serious referee.","headline":"The multimodal pest-detection results are likely inflated by label leakage through the per-species text input; the datasets and architecture are new but the main empirical claim isn't supported by the reported protocol.","tokens_in":11828,"tokens_out":5125,"would_cite":false,"duration_ms":60257,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A pest detector that fuses super-resolved images with species text descriptions claims to beat state-of-the-art unimodal detectors on IP102-derived benchmarks.","keywords":["pest detection","cross-modal fusion","image-text fusion","super-resolution","IP102 dataset","object detection","multi-modal dataset","data augmentation"],"falsifier":"Run MSFNet-CPD on CTIP102 with the text channel removed, or with text descriptions swapped randomly between species; the paper's own ablation predicts mAP falls from 46.06% to about 35.6%, which would erase most of the claimed advantage over YOLOv9 (42.32%) on the same data.","tokens_in":10787,"feed_emoji":"🐛","tokens_out":4857,"duration_ms":52658,"temperature":0.7,"pith_summary":"The paper claims that agricultural pest detection can be improved by giving the detector two extra information channels beyond the raw image: a super-resolution reconstruction of the same image, and a short text description of the target species' appearance. The proposed MSFNet-CPD feeds both the original and the reconstructed image through a multi-scale convolutional stem, converts the visual features into token sequences, and mixes them with text tokens in a Transformer-style fusion module before the YOLOv4-style detection head. On the paper's new benchmarks built from IP102, the model reports precision 82.15% and mAP 46.06% on CTIP102, outperforming the strongest unimodal baseline YOLOv9 across mAP, mAP50 (92.18% vs 81.10%), and mAP75. Ablations show that removing the text channel costs about 10 points of mAP and removing the super-resolution path costs about 11 points, so both channels are load-bearing in the reported result.","feed_headline":"Text plus super-resolution lifts pest detection past image-only models","feed_subtitle":"Fusing species descriptions and sharpened images into one Transformer beats image-only detectors on 102-class IP102 benchmarks.","key_machinery":"The load-bearing mechanism is the ITF module: a Transformer encoder (scaled dot-product attention, multi-head self-attention plus feed-forward sub-layers) that receives a concatenated sequence G built from n text tokens and 25 visual tokens at each of three scales, for both the original and super-resolved image, so the same attention operation mixes semantic class cues with fine-grained visual cues. The TIC and ITC converters map visual feature maps into and out of this token space; the LSRGAN super-resolution path supplies a second, sharper view of the same scene; and the YOLOv4 neck/head (SPP + PANet) turns the fused multi-scale features into bounding-box predictions.","core_discovery":"The central discovery the authors report is that cross-modal fusion of visual and textual features beats unimodal detection on pest benchmarks, provided the text describes the species present. The network architecture interleaves three components: LSRGAN super-resolution to repair low-quality field images; a Text-Image Converter (TIC) that condenses multi-scale visual features into 25 tokens per scale; and an Image-Text Fusion (ITF) Transformer in which text tokens and visual tokens attend to one another, with an Image-Text Converter (ITC) decoding the fused representation back to feature maps for the detection neck. The paper also contributes three datasets—STIP102 (simple text), CTIP102 (complex text), and MTIP102 (multi-target composites made by the ACIE augmentation algorithm)—and reports that more complex text yields higher precision and mAP50 than simple text.","pith_inferences":["Because the text descriptions in STIP102 and CTIP102 are constructed from the species identities, the reported mAP measures how much class-label information the model can exploit through the text channel; a deployment protocol with genuinely independent text (e.g., a field-guide entry not tailored to the image) would likely show smaller gains.","The same TIC-ITF-ITC design could transfer to any fine-grained detection task—plant diseases, medical images, insect taxonomy—where a short verbal description of the target class exists, offering a cheap way to inject semantic priors.","A fairer comparison against unimodal baselines would either give those baselines equivalent class information or evaluate MSFNet-CPD under text-only, no-text, and mismatched-text conditions; the paper reports only the no-text ablation.","The roughly 10-point gap without text suggests the visual backbone alone may underperform YOLOv9, so the headline comparison effectively contrasts a text-augmented model with an image-only model."],"forward_implications":["If text descriptions of the target species are available at inference time, pest detectors can be expected to beat image-only detectors by a wide margin on IP102-style data.","The roughly 10-point mAP drop without text indicates that almost all of the multimodal advantage depends on the semantic channel, not on the visual architecture alone.","The super-resolution path contributes about as much as the text channel, suggesting low-quality field images are a first-order bottleneck for pest detection.","The ACIE augmentation can generate labeled multi-target training images with random backgrounds and placements, which the paper reports improves robustness in complex scenes even if raw scores on MTIP102 are lower."],"supporting_citations":[{"why":"Supplies the Real-ESRGAN super-resolution module that reconstructs high-quality versions of low-resolution pest images.","marker":"[17]"},{"why":"Provides the multi-head self-attention Transformer encoder architecture that the ITF fusion module is based on.","marker":"[19]"},{"why":"Supplies the SPP/PANet neck and detection head used by MSFNet-CPD for bounding-box prediction.","marker":"[21]"},{"why":"Source of the IP102 images and of the unimodal baseline results (FRCNN, FPN, SSD300, RefineDet) reported in Table IV.","marker":"[25]"},{"why":"The YOLOv9 unimodal baseline that MSFNet-CPD claims to outperform across precision, mAP, mAP50, and mAP75.","marker":"[31]"},{"why":"Defines the bounding-box offset calculation and serves as a unimodal comparison baseline (YOLOv3).","marker":"[24]"}],"fun_headline_variants":["Text-aware fusion sharpens pest detection beyond image-only models","Cross-modal network beats image-only pest detectors on IP102","Super-res plus species text lifts pest detection accuracy","Fusing text and images improves pest detection on new benchmarks","MSFNet-CPD: mixing species text and sharpened images wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes it is a valid protocol to hand the model a text description built from the species label at inference time; if that text is removed or does not match the image's species, the reported gains largely disappear.","fun_headline_variants_meta":{"raw":{"variants":["Text-aware fusion sharpens pest detection beyond image-only models","Cross-modal network beats image-only pest detectors on IP102","Super-res plus species text lifts pest detection accuracy","Fusing text and images improves pest detection on new benchmarks","MSFNet-CPD: mixing species text and sharpened images wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1227,"prompt_tokens":993,"completion_tokens":234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":152}},"tokens_in":609,"tokens_out":234,"duration_ms":3197,"temperature":1.0,"reasoning_tokens":152,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:52:01.336367+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MSFNet-CPD on CTIP102 with the text channel removed, or with text descriptions swapped randomly between species; the paper's own ablation predicts mAP falls from 46.06% to about 35.6%, which would erase most of the claimed advantage over YOLOv9 (42.32%) on the same data.","supporting_citations":[{"cited_title":"Real-esrgan: Training real- world blind super-resolution with pure synthetic data,","cited_arxiv_id":null,"evidence_quote":"Supplies the Real-ESRGAN super-resolution module that reconstructs high-quality versions of low-resolution pest images."},{"cited_title":"Ip102: A large- scale benchmark dataset for insect pest recognition,","cited_arxiv_id":null,"evidence_quote":"Source of the IP102 images and of the unimodal baseline results (FRCNN, FPN, SSD300, RefineDet) reported in Table IV."}],"review_version":1}