{"id":"cdbf8316-8d12-4f5a-943b-8a9bdb9ab210","arxiv_id":"2505.15137","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"IC-Fusion, an infrared-centric transformer detector with a lightweight RGB backbone and gated fusion modules, achieves state-of-the-art mAP on LLVIP and competitive mAP on FLIR.","lead":"This paper presents a new way to combine visible and infrared camera images for object detection, using the infrared stream as the main source and a lightweight network for visible light. The authors report top results on one benchmark and competitive results on another, with lower computation than prior methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table III contradicts the claim that the asymmetric IR-centric setup outperforms dual ResNet-50: symmetric mAP is 46.2 vs 45.8.","rationale":"The paper's central claim is that an infrared-centric asymmetric design is highly effective and that weakening the RGB branch is not only efficient but superior. The strongest empirical support is the backbone ablation in Table III. Reading Table III carefully, the symmetric ResNet50/ResNet50 configuration achieves mAP 46.2, while the proposed ResNet18/ResNet50 achieves mAP 45.8; the text claims the asymmetric setup 'even outperforms' the dual ResNet-50 configuration. This is internally inconsistent if mAP is the primary metric, as it is used throughout Tables I-III. The only interpretation that saves the sentence is selecting mAP50 (83.9 vs 83.4), but the paper does not say that, and the abstract and conclusion make a general superiority claim. This is load-bearing because the entire architecture is motivated by the premise that IR should carry the heavy feature extraction; if symmetric fusion is actually more accurate on the primary metric, the paper's central comparative claim is unsupported and the contribution reduces to a modest efficiency trade-off. The reader's weakest_assumption concerned generalization of the wavelet motivation; our concern is more direct: a within-paper contradiction. We propose a multi-seed rerun to check whether the 0.4 mAP gap is stable. Verdict remains CONDITIONAL: the paper should correct the claim and either report the symmetric baseline as higher-accuracy or explicitly justify using mAP50 as the selection criterion, with error bars for the differences.","tokens_in":7849,"tokens_out":3722,"duration_ms":31763,"concrete_test":"Rerun the two backbone configurations in Table III (ResNet18/RGB + ResNet50/IR versus ResNet50/RGB + ResNet50/IR) using the released code, a fixed seed protocol, and at least three seeds; report mean and standard deviation of mAP on FLIR-aligned. If ResNet50/50 mAP remains greater than or equal to ResNet18/50 mAP, the sentence claiming the asymmetric setup 'even outperforms' the dual ResNet-50 configuration must be corrected, and the IR-centric superiority thesis should be re-evaluated as an efficiency trade-off rather than an accuracy improvement.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section III.B states that the asymmetric setup 'even outperforms the dual ResNet-50 configuration,' but Table III reports ResNet18(RGB)+ResNet50(IR) at mAP 45.8 and ResNet50(RGB)+ResNet50(IR) at mAP 46.2. The symmetric baseline is 0.4 mAP higher on the primary metric used throughout Tables I-III; only mAP50 favors the asymmetric setup (83.9 vs 83.4), and mAP75 is essentially tied (42.6 vs 42.4). This is an internal inconsistency, not a disagreement with external consensus. It is load-bearing because the entire architecture is motivated by the premise that IR should carry heavy feature extraction while RGB is deliberately weakened; if symmetric fusion is actually more accurate on the primary metric, the central comparative claim is unsupported. The result may still be defensible as an efficiency-performance trade-off, but the paper's text overstates what the table shows. The reader's concern about generalizing the wavelet motivation is related, but the more direct problem is that the paper's own data already fail to support the claimed superiority of the asymmetric design.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"IC-Fusion is a multispectral object detector built on the RT-DETR transformer framework with deliberately asymmetric backbones: a lightweight ResNet-18 for the RGB modality and a deeper ResNet-50 for the IR modality. The design is motivated by a wavelet decomposition (Fig. 1) and by unimodal FLIR results (Table I) suggesting that IR carries the high-frequency structural cues most critical for detection. A fusion module comprising Multi-Scale Feature Distillation (MSFD) for the RGB stream, a Cross-Modal Channel Shuffle Gate (CCSG), a Cross-Modal Large Kernel Gate (CLKG), and a Channel Shuffle Projection (CSP) is inserted before the transformer encoder. The paper reports FLIR-aligned mAP of 46.1 and LLVIP mAP of 70.3, claims state-of-the-art performance on LLVIP, and presents ablations on backbone asymmetry (Table III) and on each proposed component (Table IV), along with parameter/MAC comparisons (Table V).","tokens_in":8059,"tokens_out":9793,"duration_ms":74575,"significance":"The paper has several concrete strengths: it releases code, evaluates on the two standard benchmarks in the area (FLIR-aligned and LLVIP), and the component ablation in Table IV is directionally consistent, with each added module improving the previous configuration. The central design hypothesis — that a deliberately weakened RGB branch combined with a deep IR branch can match or exceed symmetric dual-ResNet-50 fusion — is falsifiable and worth testing. However, the significance is currently bounded by three factors. First, the reported advantages over GM-DETR are extremely thin (70.3 vs 70.2 mAP on LLVIP; 46.1 vs 45.8 mAP on FLIR) with no error bars or multiple-seed reporting. Second, the motivating premise that IR is the dominant modality is contradicted by the paper's own unimodal LLVIP numbers, where RGB RT-DETR achieves 67.9 mAP versus 54.2 mAP for IR. Third, the claim in Section III.B that the asymmetric setup 'even outperforms the dual ResNet-50 configuration' is not supported by Table III on the primary metric (46.2 vs 45.8 mAP).","major_comments":[{"comment":"The same configuration — ResNet-18 RGB backbone, ResNet-50 IR backbone, and the full MSFD+CCSG+CLKG fusion module — is reported with different performance numbers on FLIR-aligned across tables. Table II lists mAP50 83.3 / mAP75 42.6 / mAP 46.1 for IC-Fusion, while Table III (ResNet18/ResNet50 row) and Table IV (final ablation row) list 83.9 / 42.6 / 45.8 for the same model. Because these tables describe the identical architecture, the 0.6-point mAP50 and 0.3-point mAP discrepancies must be explained (e.g., different seeds, checkpoints, or evaluation settings); without such an explanation the headline FLIR result and the ablation baseline are mutually inconsistent, and the reader cannot tell which numbers are the definitive ones.","section":"Tables II-IV, Section III.B"},{"comment":"The sentence 'our asymmetric setup even outperforms the dual ResNet-50 configuration despite having fewer parameters and reduced computational cost' is not supported by Table III on the paper's primary metric. Table III reports mAP 46.2 for ResNet50+ResNet50 versus 45.8 for ResNet18+ResNet50; the asymmetric configuration leads only on mAP50 (83.9 vs 83.4) and ties on mAP75 (42.6 vs 42.4). Since the experimental setup defines mAP averaged over IoU 0.50–0.95 as the headline measure, the symmetric dual-ResNet-50 setup is more accurate on the primary metric. The claim should be restated as an efficiency-accuracy trade-off, or the text revised, because as written the central comparative claim for the IR-centric asymmetric design is contradicted by the paper's own data.","section":"Section III.B, Table III"},{"comment":"The motivating premise that IR images contain 'structurally rich high-frequency information critical for object detection' is supported only on FLIR-aligned and is contradicted on LLVIP, where RT-DETR in RGB mode reaches 67.9 mAP versus 54.2 mAP in IR mode (Table II). The wavelet evidence in Fig. 1 is a single qualitative example with no quantitative link between sub-band energy and detection accuracy. Because the architecture deliberately assigns the deeper backbone to IR and the shallower backbone to RGB, the paper should either restrict the motivating claim to FLIR-like conditions or provide a quantitative analysis (e.g., sub-band energy statistics across both datasets, or a backbone-depth-controlled experiment) explaining why an IR-centric design is justified on a benchmark where unimodal IR is substantially weaker.","section":"Section I, Tables I-II"},{"comment":"No error bars, seeds, or number of runs are reported anywhere in the experiments, and the decisive margins are very small: 70.3 vs 70.2 mAP on LLVIP against GM-DETR, 46.1 vs 45.8 mAP on FLIR, and ablation increments of 0.4–0.7 mAP in Table IV. For a DETR-style training pipeline with 60 epochs and a batch size of 8, these differences are likely within run-to-run variance. The authors should report the mean and standard deviation over at least three seeds and state explicitly whether the state-of-the-art claim and the component-ablation claims survive that comparison.","section":"Section III, Tables II-IV"}],"minor_comments":[{"comment":"The mAP75 entry for Yolov5 (IR) is a dash even though mAP75 is reported for all other rows in the table; either provide the value or mark the column as not evaluated for that model.","section":"Table I"},{"comment":"The reference title contains a typo: 'Generalized muiltispectral' should read 'Generalized Multispectral'.","section":"Reference [5]"},{"comment":"The two-stage training strategy is described in a single sentence; it would help to specify concretely what 'modality-isolated pretraining' involves (epochs, data splits, whether the fusion modules are trained in the second stage from scratch) and to state whether the same protocol is used for the ablations in Tables III and IV.","section":"Section III.A"},{"comment":"The efficiency comparison reports parameters and MACs but not wall-clock latency; given the paper's 'lightweight' and 'efficient' claims, an FPS measurement on a common GPU would strengthen the comparison, and the input resolution used for the MACs computation (presumably 640x640) should be stated.","section":"Section III.B, Table V"},{"comment":"The statement that IC-Fusion 'consistently outperforms prior fusion methods' is stronger than Table II supports, since GM-DETR still leads on FLIR mAP50 (83.9 vs 83.3) and ties on mAP75 (42.6 vs 42.6); the claim should be qualified to the metrics and datasets where the advantage actually holds.","section":"Section III.B"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue for the editor is the cross-table discrepancy for the identical configuration (Table II vs Tables III and IV), which suggests the headline numbers and the ablation numbers may come from different training protocols, seeds, or evaluation settings; this needs to be resolved before any revision can be assessed. The margins over GM-DETR are within typical noise for this training setup, so multi-seed reporting is essential. The manuscript is formatted as a journal letter, though the contribution is at the level of a solid conference paper; the authors may wish to align the presentation with the target venue. I saw no citation-pattern concerns beyond the heavy reliance on the GM-DETR training recipe, which is properly acknowledged."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real, reproducible architecture paper with new modules and public code, but the central 'IR-centric beats symmetric' claim is contradicted by the paper's own Table III. That's the first thing to know. Second: the reported margins over GM-DETR are tiny, and there are no error bars, so the empirical edge is fragile.\n\nWhat's actually new: the combination of asymmetric backbones (ResNet-18 RGB / ResNet-50 IR) with the MSFD, CCSG, CLKG, CSP fusion blocks is not in prior work. The ablations in Table IV are clean and show each component adds a little. The code is public, and the LLVIP result (70.3 mAP) is the best reported there, though by 0.1 over GM-DETR. That's a legitimate contribution to the multispectral detection subfield.\n\nThe soft spots: the paper overclaims. Section III.B says the asymmetric setup 'even outperforms the dual ResNet-50 configuration,' but Table III shows the opposite: ResNet18/ResNet50 gets 45.8 mAP, ResNet50/ResNet50 gets 46.2. Only mAP50 favors the asymmetric. So the load-bearing comparative claim is unsupported by the authors' own data. There's also a second inconsistency: Table II reports the final model at 46.1 mAP while the full ablation row in Table IV says 45.8. These aren't explained. On top of that, the wavelet-based motivation (Fig. 1) is a single qualitative example, and no error bars or multiple runs back the headline numbers. The margins over GM-DETR on both benchmarks (0.3 and 0.1) are within what could easily be seed variation.\n\nThe paper is still worth refereeing: the architecture is non-trivial, the code is out, and the efficiency story (56M params vs 66M, 88G MACs vs 104G for RT-DETR multi) is real. But a referee should demand that the authors correct the Table III misstatement, reconcile Tables II and IV, and provide variance estimates. Without those, the central claim doesn't stand.\n\nWho's this for: people working on visible-infrared fusion or efficient DETR variants. It's not a breakthrough, but it's a usable baseline. I'd send it to peer review with a 'major revision' expectation.","headline":"Real architecture and public code, but the paper's central 'IR-centric beats symmetric' claim is contradicted by its own Table III, so it needs major revision before the empirical story is credible.","tokens_in":8560,"tokens_out":6441,"would_cite":false,"duration_ms":49216,"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":"This paper claims that infrared should be the primary modality in multispectral object detection, with a lightweight RGB branch providing supporting semantic context, and that this asymmetric design reaches state-of-the-art accuracy at…","keywords":["multispectral object detection","infrared-centric fusion","detection transformer","cross-modal gating","feature distillation","RT-DETR","LLVIP","FLIR"],"falsifier":"Run IC-Fusion against a symmetric ResNet-50/ResNet-50 fusion on a multispectral benchmark where visible images carry most of the boundary detail, such as a daytime high-texture scene set, or compute wavelet sub-band energies across FLIR and LLVIP and correlate them with per-image mAP; if the symmetric model wins or RGB-heavy high-frequency images do not favor the IR-centric design, the central claim fails.","tokens_in":7645,"feed_emoji":"🌡️","tokens_out":5929,"duration_ms":51297,"temperature":0.7,"pith_summary":"This paper tries to establish that infrared is the modality that should carry most of the detection load in multispectral object detection, with RGB playing a supporting role. The authors support this with wavelet decompositions showing that IR sub-bands carry sharp object boundaries, and with unimodal FLIR results where IR-only detectors beat RGB-only detectors. On that basis they build IC-Fusion, a detection transformer pairing a lightweight ResNet-18 RGB stream with a deeper ResNet-50 IR stream, fused through gated cross-modal modules. IC-Fusion reports 70.3 mAP on LLVIP and 46.1 mAP on FLIR-aligned, matching or beating symmetric ResNet-50/ResNet-50 fusion while using fewer parameters and multiply-accumulate operations. A reader should care because the claim, if true, redirects multispectral detector design toward asymmetric, infrared-prioritized architectures rather than equal-weight fusion.","feed_headline":"IR-centric fusion hits 70.3 mAP on LLVIP, beats symmetric nets","feed_subtitle":"A lightweight RGB stream plus a deep IR stream outperforms symmetric ResNet-50/ResNet-50 fusion at lower cost.","key_machinery":"The central object is IC-Fusion, an infrared-centric multispectral detection transformer built on the RT-DETR detector. Its load-bearing asymmetry is that the RGB stream uses a compact ResNet-18 backbone while the IR stream uses a deeper ResNet-50 backbone. The fusion machinery is a cascade of four modules: Multi-Scale Feature Distillation (MSFD), which enriches RGB features with multi-branch depthwise convolutions; Cross-Modal Channel Shuffle Gate (CCSG), which reorganizes channels across modalities and gates them with GELU; Cross-Modal Large Kernel Gate (CLKG), which applies large-kernel depthwise convolutions as a spatial gate to align long-range structure; and Channel Shuffle Projection (CSP), which distills the concatenated multi-scale features with grouped pointwise convolutions. Together they let the model emphasize informative cross-modal features and suppress redundant ones, enabling the asymmetric backbone split to work.","core_discovery":"The paper's central claim is that IR images are not merely a fallback for low-light conditions but the structurally dominant modality: their high-frequency sub-bands contain the boundaries and contours that localization depends on, while RGB mainly provides low-frequency semantic context. From this the paper derives a design principle, allocate more capacity to IR and keep the RGB branch deliberately lightweight, and instantiates it in IC-Fusion with ResNet-50 for IR, ResNet-18 for RGB, and a fusion module where MSFD enriches RGB features, CCSG and CLKG gate cross-modal features, and CSP projects the result. The reported outcome is a state-of-the-art 70.3 mAP on LLVIP and 46.1 mAP on FLIR-aligned, with an ablation showing that the asymmetric backbone choice beats the reversed assignment and the symmetric ResNet-50/ResNet-50 configuration while using fewer parameters.","pith_inferences":["Going beyond the paper, the wavelet evidence is a single illustrative decomposition rather than a statistical link between sub-band energy and detection accuracy; measuring that link across FLIR and LLVIP would tell whether the IR-centric premise is quantitative or qualitative.","If IR-centricity generalizes, the likely failure boundary is daytime, high-texture scenes where visible images carry most boundary detail; in those conditions adaptive routing or symmetric fusion could beat a fixed IR-prioritized design.","The lightweight gating primitives, channel shuffle plus GELU gating, could plausibly transfer to other cross-modal tasks such as semantic segmentation or visual-infrared tracking, not just object detection.","A testable extension is to apply the same asymmetric backbone split to other transformer detectors and see whether the LLVIP and FLIR margins persist independently of RT-DETR's specific encoder."],"forward_implications":["If IR is structurally dominant, multispectral detectors can shift compute from RGB to IR without losing accuracy, as shown by the ResNet-18/ResNet-50 configuration beating ResNet-50/ResNet-50 at lower cost.","Each component in the fusion cascade contributes measurably on FLIR-aligned, so the full MSFD-CCSG-CLKG sequence, not any single gate, is what carries the gain.","Since unimodal IR already outperforms unimodal RGB on FLIR across backbone types, IR-centric fusion should also hold on other benchmarks where thermal contours dominate over visible texture.","Because the method is built on RT-DETR, the infrared-centric fusion principle can be transferred to other DETR-family detectors without redesigning the encoder-decoder.","The two-stage training strategy, pretraining on modality-isolated data before fusion training, is part of what makes the asymmetric design work on the limited aligned data."],"supporting_citations":[{"why":"RT-DETR is the base detector and efficient encoder into which the fusion module is plugged; its unimodal IR and RGB numbers motivate the asymmetry.","marker":"[9]"},{"why":"Provides the aligned version of the FLIR benchmark used for all FLIR experiments and ablations.","marker":"[8]"},{"why":"LLVIP is the low-light benchmark where IC-Fusion reports the state-of-the-art 70.3 mAP.","marker":"[21]"},{"why":"FLIR is the thermal dataset whose aligned pairs, curated in [8], supply the second benchmark and the unimodal comparisons.","marker":"[20]"},{"why":"GM-DETR is the strongest symmetric multispectral baseline on both datasets and the source of the two-stage training strategy used here.","marker":"[5]"}],"fun_headline_variants":["Infrared-centric fusion beats symmetric networks at lower cost","Deep IR, shallow RGB: the fusion recipe that wins","IR-first fusion: higher mAP, fewer parameters","Make infrared the primary sensor in multispectral fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that infrared images generally hold the high-frequency boundary structure that object detection needs, so the deep network should go to IR and the RGB stream can be lightweight; this is supported by one wavelet example and by unimodal FLIR results, but it is assumed to hold across other datasets and lighting conditions.","fun_headline_variants_meta":{"raw":{"variants":["Infrared-centric fusion beats symmetric networks at lower cost","Deep IR, shallow RGB: the fusion recipe that wins","IR-first fusion: higher mAP, fewer parameters","Make infrared the primary sensor in multispectral fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001544,"raw_usage":{"total_tokens":6167,"prompt_tokens":931,"completion_tokens":5236,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":5173}},"tokens_in":547,"tokens_out":5236,"duration_ms":37291,"temperature":1.0,"reasoning_tokens":5173,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:22:23.487293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run IC-Fusion against a symmetric ResNet-50/ResNet-50 fusion on a multispectral benchmark where visible images carry most of the boundary detail, such as a daytime high-texture scene set, or compute wavelet sub-band energies across FLIR and LLVIP and correlate them with per-image mAP; if the symmetric model wins or RGB-heavy high-frequency images do not favor the IR-centric design, the central claim fails.","supporting_citations":[{"cited_title":"Free teledyne flir thermal dataset for algorithm training,","cited_arxiv_id":null,"evidence_quote":"FLIR is the thermal dataset whose aligned pairs, curated in [8], supply the second benchmark and the unimodal comparisons."},{"cited_title":"Gm-detr: Generalized muiltispectral detection transformer with efficient fusion encoder for visible-infrared detection,","cited_arxiv_id":null,"evidence_quote":"GM-DETR is the strongest symmetric multispectral baseline on both datasets and the source of the two-stage training strategy used here."}],"review_version":1}