{"id":"acf197b7-76df-46fe-b2e8-ec0ec7bd2d84","arxiv_id":"2607.03795","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"QualGate-regulated RGB guidance during training produces efficient IR-only and dual-modal detectors that match or beat equal-fusion baselines under low light and adverse weather.","lead":"InfraNet trains infrared object detectors with RGB used only as quality-gated auxiliary guidance, then deploys either IR-only or dual-modal models. This yields accurate night and adverse-weather detection without forcing RGB at test time, useful for cars, drones, and surveillance.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-noted hand-designed scalar gain.","rationale":"The paper is a clean systems contribution: IR-primary pathway + QualGate (Eqs. 1–4) that jointly suppresses RGB and amplifies IR, two exportable architectures, and consistent gains on LLVIP/FLIR/M3FD/DroneVehicle (Tables 1–4) plus the design ablations (Tables 5–7). The IR-only export path is practical and correctly zero-overhead at inference. The scalar-q + fixed-clip design is the weakest modeling assumption, exactly as the reader flagged, yet the paper already demonstrates necessity of both gates and task-oriented behavior of q under fog/brightness. That is enough for the performance claim; it does not require external quality labels or multi-channel reliability maps. Missing multi-seed stats and code are real but secondary for a CV empirical paper and already justify the reader's CONDITIONAL. No stronger load-bearing flaw (e.g., unfair baselines, train/test leakage, or contradictory ablations) appears. Therefore leave the verdict unchanged.","tokens_in":17124,"tokens_out":559,"duration_ms":5486,"concrete_test":"Re-train the YOLOv8 InfraNet-IR configuration on LLVIP with α(q)=clip(c-q;1.0,c) for c∈{1.2,1.5,2.0} (and optionally a learned free α) under identical seeds/protocol; if mAP50-95 stays within ~0.5–1 point of the reported 68.8 across the range, the hand-chosen constant is non-critical and the claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (scalar q from GAP(A_in) plus the hand-chosen clip(1.5-q;1.0,1.5) IR gain) is the softest design choice, but it is not load-bearing for the central claim. The claim is empirical: QualGate-trained InfraNet yields competitive/strong mAP on four RGB-IR benchmarks in both IR-only and RGB-IR modes, with zero extra IR-only deployment cost. Tables 5–6 and the w_aux ablation already show that removing either suppression or amplification hurts, that q responds to controlled RGB degradation, and that the asymmetric loss works. The constant 1.5 is explicitly non-unique; richer reliability modeling is not required for the reported numbers to hold. No internal inconsistency or missing control that would reverse the tables is present.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes InfraNet, an IR-centric multi-modal detection framework that treats RGB as reliability-controlled auxiliary guidance rather than a co-equal modality. Its core module, QualGate (§3.1, Eqs. 1–4), predicts a scalar task-oriented reliability score q from GAP of the auxiliary features and uses it both to suppress RGB contributions and to amplify multi-scale IR features via a clipped inverse gain α(q). Two independently trained variants are defined: InfraNet-IR (RGB used only at training; pure IR backbone–neck–head at inference) and InfraNet-RGB-IR (dual-stream at both train and test). Experiments on LLVIP, FLIR-Aligned, M3FD and DroneVehicle across ResNet50 and YOLO-family backbones report competitive or superior mAP/mAP50 relative to recent fusion methods (WaveMamba, CFT, etc.), with IR-only models retaining the parameter/latency footprint of a single-stream detector. Ablations isolate RGB suppression vs. IR amplification (Table 5), controlled RGB degradation behavior of q (Table 6), and the auxiliary loss weight w_aux (Table 7).","tokens_in":17370,"tokens_out":1258,"duration_ms":10185,"significance":"If the reported numbers hold under independent re-implementation, the work supplies a practical privileged-guidance recipe that decouples training-time multi-modal supervision from test-time modality availability—an important deployment axis for thermal perception under night, fog and UAV conditions. Strengths that raise the contribution above a pure engineering tweak include: (i) consistent gains across four public RGB–IR benchmarks and multiple detector families, (ii) explicit zero-overhead IR-only export, (iii) controlled degradation sweeps that show q responds to RGB corruption without external quality labels, and (iv) component ablations that separate suppression from amplification. The design is simple enough to be portable; the main novelty is the asymmetric training/inference contract rather than a new backbone.","major_comments":[{"comment":"§5.2 / Tables 1–4 report only single-run point estimates with no standard deviations, multiple seeds, or statistical tests. Given that several claimed margins over WaveMamba are modest (e.g., LLVIP YOLOv8 mAP 70.5 vs 66.0; FLIR mAP gaps of a few points), the ranking of “strong or competitive” cannot be assessed for stability. At minimum the authors should re-run the primary YOLO/ResNet configurations with 3 seeds and report mean±std for mAP50 and mAP, or supply a clear statement that only single-run numbers are available and why.","section":null},{"comment":"§3.1 Eqs. (2)–(4) and the accompanying text leave the IR amplification schedule under-specified for reproducibility. α(q)=clip(1.5−q;1.0,1.5) is acknowledged as non-unique, yet the paper does not report a sensitivity sweep over the upper bound (or alternative functional forms) in the main text; the supplementary is only alluded to. Because the dual-path design is presented as load-bearing (Table 5), a short main-text ablation of the clip range (or a learned α) is needed so that readers can judge how tightly the reported gains depend on the hand-chosen constant.","section":null},{"comment":"§5.2 Table 4 (M3FD Day/Overcast/Night/Challenge) is labeled a “diagnostic comparison” and notes that many recent methods do not report the same subset protocol, yet the table still ranks InfraNet first by large margins against older fusion baselines. Either the authors should re-evaluate the strongest recent competitors (WaveMamba, RSDet, etc.) under the identical subset split, or they should move Table 4 to the appendix and avoid presenting it as a primary SOTA comparison. As written, the subset numbers risk overstating the method’s advantage.","section":null}],"minor_comments":[{"comment":"Fig. 1 caption and surrounding text claim that CFT/FusionMamba/WaveMamba “introduce additional noise” relative to pure IR; the visualization is qualitative channel-sum heatmaps. A quantitative noise or SNR proxy (or at least a clearer description of how the maps were normalized) would make the motivation more rigorous.","section":null},{"comment":"§3.2 / Fig. 2: the description of which features ({A3,A4,A5} vs {P3,P4,P5}) feed the detection head during training for InfraNet-IR is slightly ambiguous; a one-sentence clarification of the dual-head loss paths would help re-implementation.","section":null},{"comment":"Notation: Ain is used both for “RGB feature or fused feature” (§3.1) and later as the auxiliary input at each site; a short glossary or consistent subscripting would reduce confusion.","section":null},{"comment":"Related work (§2.2) cites HalluciDet, CRLM and CAGF; Table 5 already replaces QualGate with those modules, which is good, but a one-line discussion of why a scalar q is preferred over their denser reliability maps would strengthen the positioning.","section":null},{"comment":"Minor typos / style: “M3FD” vs “M$^3$FD” inconsistency; “compensateIRfeaturesduringcross-modaltraining” spacing artifacts in the abstract; “tip aggregation” in the DroneVehicle discussion appears to be a typo for “feature aggregation”.","section":null},{"comment":"Inference-time numbers in Table 2 (ms) are useful; stating the batch size and whether TensorRT/ONNX was used would improve comparability.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is a solid empirical systems contribution rather than a conceptual breakthrough; the scalar-q + clipped-gain design is simple and the weakest modeling assumption is not load-bearing for the tables. Fit for a solid CV / multi-modal perception venue is good after the statistical and reproducibility points above are addressed. No integrity or citation-pattern concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a practical systems paper, not a theory one. The useful piece is the asymmetric setup: RGB is privileged guidance only during training, QualGate regulates it with a learned scalar q that both suppresses bad RGB and amplifies IR, and InfraNet-IR then drops the RGB branch and fusion modules at inference with zero extra cost. That export path is the real product contribution; dual-modal InfraNet-RGB-IR is the higher-capacity option when both sensors are available.\n\nWhat they do well: four standard RGB-IR benchmarks (LLVIP, FLIR-Aligned, M3FD, DroneVehicle), multiple YOLO/ResNet backbones, and ablations that actually isolate the design. Table 5 shows that swapping QualGate for HalluciDet-style pseudo-RGB, CAGF, or CRLM loses ground, and that killing either the RGB-suppression or IR-amplification path hurts. Table 6 shows q moves under controlled fog/brightness on fixed IR, and the asymmetric w_aux=0.25 loss is checked. Night/challenge subsets and feature visualizations support the IR-centric story. Citations to LUPI, HalluciDet, CRLM, CAGF are present and the comparisons are fair enough. No circular math; pure empirical claims.\n\nSoft spots, in proportion: the IR gain α(q)=clip(1.5−q;1.0,1.5) is hand-chosen (they say so), q is only a GAP+MLP scalar, and there are no multi-seed error bars or released code. Those are real but not load-bearing for the reported tables. Free parameters (w_aux, clip range, MLP size) are few and ablated. This is engineering progress for night/UAV/surveillance detection, not a reorg of multi-modal CV.\n\nWho it is for: people shipping IR detectors who want RGB help at train time without paying for it at deploy time. Worth a serious referee. I would send it to peer review; expect requests for seeds, code, and a slightly richer reliability control, not a rewrite of the core idea.","headline":"Clean IR-centric training/export pattern with solid multi-benchmark numbers; the scalar QualGate is simple but the ablations hold.","tokens_in":17945,"tokens_out":524,"would_cite":true,"duration_ms":4896,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"InfraNet trains infrared detectors with quality-gated RGB help, then drops RGB for free IR-only inference.","keywords":["infrared object detection","quality-aware fusion","modality-asymmetric learning","RGB-IR","privileged information","QualGate","IR-only deployment"],"falsifier":"Train identical InfraNet models with and without QualGate on a night-only split where RGB is heavily degraded; if the gated IR-only model no longer outperforms a pure-IR baseline or a naïve equal-weight fusion baseline on that split, the quality-control claim fails.","tokens_in":18054,"feed_emoji":"🕶️","tokens_out":791,"duration_ms":7838,"temperature":0.7,"pith_summary":"Standard RGB–infrared detectors fuse both streams equally and still need both cameras at test time, so a dark or foggy RGB image can hurt the very system it is meant to help. InfraNet reverses the priority: the infrared branch is the primary path that produces the final detections, while RGB is treated only as optional training-time guidance. A lightweight gate called QualGate predicts a single reliability score from the RGB features; when that score is low it down-weights the RGB contribution and simultaneously boosts the infrared features so the network never learns from contaminated cues. After training, the entire RGB branch and all gates can be stripped away, leaving a pure infrared detector whose accuracy still reflects the privileged RGB supervision. On four standard benchmarks the resulting IR-only models match or beat heavier dual-modal systems while running at the same cost as ordinary single-modal detectors.","feed_headline":"RGB trains IR detectors, then is thrown away at test time","feed_subtitle":"A quality gate lets infrared models keep multi-modal gains without dual-camera cost at inference.","key_machinery":"QualGate: a multi-scale fusion module that predicts a scalar reliability score q from global-average-pooled RGB features, multiplies the RGB stream by q (suppression) and multiplies the IR stream by a clipped inverse gain α(q) (compensation), thereby shaping IR-centered representations without external quality labels.","core_discovery":"An asymmetric, IR-centric training scheme that uses RGB solely as reliability-controlled auxiliary supervision can produce infrared detectors that are both more accurate under adverse lighting and free of any dual-modal cost at inference. The same architecture can keep the RGB stream when it is available, giving a higher-capacity dual-modal detector under the same quality-aware regulation.","pith_inferences":["The same scalar-gate idea could be applied to other privileged-modality pairs (depth, event cameras, radar) where one stream is cheap at inference and the other is only intermittently reliable.","Because the gate never requires explicit quality labels, the method may transfer to unlabeled or weakly paired multi-modal corpora that lack night/day annotations.","If the learned score q correlates with scene illumination, it could later serve as a cheap online sensor-health signal for adaptive fusion policies."],"forward_implications":["IR-only detectors can inherit the accuracy gains of multi-modal training without carrying multi-modal compute or sensor cost at deployment.","When both cameras remain available, the same quality gate can be left active, yielding a higher-capacity dual-modal detector under one training recipe.","Negative transfer from degraded RGB can be treated as a controllable training signal rather than an inevitable fusion risk.","Any multi-scale detector family can host the same asymmetric QualGate pattern, so the idea is not tied to a single backbone."],"fun_headline_variants":["RGB quality gates train IR detectors that drop cameras at test","Asymmetric InfraNet uses RGB only as training supervisor for IR","QualGate lets IR detectors keep multi-modal gains without dual cost","RGB assists IR via reliability control then vanishes at inference","IR-centric training yields accurate detectors free of dual sensors"],"cache_read_input_tokens":128,"weakest_assumption_plain":"A single scalar score computed from a global pool of RGB features, together with a fixed hand-chosen amplification schedule, is enough to decide when RGB is helpful and when it should be ignored.","fun_headline_variants_meta":{"raw":{"variants":["RGB quality gates train IR detectors that drop cameras at test","Asymmetric InfraNet uses RGB only as training supervisor for IR","QualGate lets IR detectors keep multi-modal gains without dual cost","RGB assists IR via reliability control then vanishes at inference","IR-centric training yields accurate detectors free of dual sensors"]},"model":"grok-4.5","effort":"low","cost_usd":0.006348,"raw_usage":{"total_tokens":1648,"prompt_tokens":789,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":63480000,"prompt_tokens_details":{"text_tokens":789,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":774,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":789,"tokens_out":85,"duration_ms":6112,"temperature":1.0,"reasoning_tokens":774,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T23:53:55.337906+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train identical InfraNet models with and without QualGate on a night-only split where RGB is heavily degraded; if the gated IR-only model no longer outperforms a pure-IR baseline or a naïve equal-weight fusion baseline on that split, the quality-control claim fails.","supporting_citations":[],"review_version":1}