{"id":"10b2e2fc-c31a-4f13-b31b-6cf0fbaa12ac","arxiv_id":"2507.18173","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"WaveMamba fuses RGB and infrared features in the wavelet domain and reports an average mAP gain of about 4.5 points over prior methods on four public benchmarks.","lead":"WaveMamba fuses visible and infrared images for object detection by decomposing both into high- and low-frequency parts with the Haar wavelet, then combining low-frequency content with Mamba-based blocks and high-frequency edges by keeping the larger peaks. It reports the highest detection accuracy on four public RGB-infrared benchmarks, with an average gain of about 4.5 percentage points over prior methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 4.5% SOTA gain is measured against literature-reported baselines, and the paper's own protocol baseline differs by several mAP points from those numbers, so an apples-to-apples re-run is needed before the headline claim can be accepted.","rationale":"Reader's weakest_assumption is the same as mine: the headline number depends on copied baseline results. My stress-test adds the observation that the paper's own Table 7 baseline is far below the SOTA baselines it compares against on M3FD, which makes protocol mismatch concrete rather than hypothetical. I do not see a more basic flaw: the DWT/Mamba construction is described in enough detail to be plausible, the ablations are internally consistent, and the claimed average is arithmetically consistent with Tables 1-4. The concern is about external validity of the comparison, not the soundness of the architecture. Because the paper is otherwise a normal empirical study and the issue can be settled by re-running baselines, the appropriate verdict is still CONDITIONAL rather than ACCEPT or REJECT. No change to the reader's verdict is needed.","tokens_in":20281,"tokens_out":6733,"duration_ms":69952,"concrete_test":"Using released implementations where available, re-run the top two published baselines per benchmark (e.g., MRD-YOLO and MMFN on M3FD; IV-YOLO and CMA on DroneVehicle; MMSANet and FAWDet on LLVIP; ESSFN and CrossFormer on FLIR-Aligned) under WaveMamba's exact protocol: the same official split, 640x640 input, 250 epochs, batch size 16, SGD momentum 0.937, weight decay 0.0005, and default YOLOv8 augmentation, then recompute mAP and the average per-dataset gap. If the corrected average remains at least 4.5% and every per-benchmark gap remains positive, the concern is resolved; if the average falls below 3.5% or any gap reverses, the headline claim should be revised. For baselines without public code, the authors should at least state which settings were used for the cited numbers and re-run the ones with code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is the average mAP improvement of 4.5% over the second-best SOTA method, as stated in the abstract and §1 and quantified in Tables 1-4. On its face the arithmetic checks out: the four per-dataset mAP gaps over the best published baseline are approximately 5.1 (M3FD), 3.7 (DroneVehicle), 3.5 (LLVIP), and 5.8 (FLIR-Aligned), averaging 4.5. The load-bearing assumption is that those baseline numbers were produced under conditions comparable to WaveMamba's training protocol. Section 4.1 specifies only WaveMamba's own settings (split, 640x640, 250 epochs, batch 16, SGD, augmentation defaults); it nowhere states that the compared methods were re-run under those settings. Internal evidence suggests protocol sensitivity is real: Table 7 shows that under the authors' own pipeline, a plain average-fusion dual-stream YOLOv8 baseline with the original head reaches only 83.2 mAP50 / 55.1 mAP on M3FD, while the cited SOTA numbers for MMFN and MRD-YOLO are 86.2/57.4 and 86.6/59.3. If the baselines were re-run in the authors' framework, the gaps in Tables 1-4 could shift by several points in either direction, and the 4.5% average might not survive. The paper also reports no error bars and does not release code, so the differences cannot be attributed to method versus protocol. Thus the central numerical claim is under-specified and potentially over-stated, though not internally inconsistent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"WaveMamba is a dual-stream RGB-infrared object detector built on a YOLOv8 framework. The authors apply a Haar discrete wavelet transform (DWT) to backbone features, fuse the low-frequency sub-bands with a Mamba-based Low-frequency Mamba Fusion Block (channel swapping plus a gated-attention deep fusion module), fuse high-frequency sub-bands with an 'absolute maximum' strategy, and use an inverse DWT in a modified detection head. Experiments are reported on M3FD, DroneVehicle, LLVIP, and FLIR-Aligned in the main text, with additional results on VEDAI and KAIST in the supplement. The paper claims an average mAP improvement of 4.5% over the second-best state-of-the-art method. Ablations on M3FD support the contribution of the improved head, the SFM and DFM modules, and the HFE/LMFB combination.","tokens_in":20660,"tokens_out":8636,"duration_ms":89123,"significance":"The frequency-complementarity motivation is clean, and the proposed modules are simple enough that the approach could be practically useful if the reported gains are real. The paper covers six datasets and three backbones, and the ablation tables show that each module contributes to the final result. However, the headline numerical claim is not yet established at the level expected for a journal publication: the comparisons are against literature-reported baseline numbers without a controlled re-run under one protocol, the final architecture was selected using the same M3FD benchmark that appears in the headline comparison, and no code or error bars are provided. Because the central claim is purely empirical, these protocol issues determine whether the 4.5% figure is meaningful.","major_comments":[{"comment":"The 4.5% average gain is computed by comparing WaveMamba's numbers with baseline numbers taken from prior publications, but the manuscript does not state that those methods were re-run under WaveMamba's protocol (640x640 input, 250 epochs, batch size 16, SGD, default YOLOv8 augmentation). The sensitivity of the comparison to protocol is visible in Table 7, where the authors' own average-fusion dual-stream YOLOv8 baseline reaches only 83.2 mAP50 / 55.1 mAP on M3FD, while the cited MRD-YOLO value on the same YOLOv8 backbone is 86.6 / 59.3. If the prior baselines were trained with different splits, image sizes, epochs, or augmentation, the per-dataset gaps in Tables 1-4, and therefore the claimed 4.5% average, could change materially. A controlled re-run of at least the closest competitors under one protocol is needed before the headline claim can be accepted.","section":"4.1; Tables 1-4; Table 7"},{"comment":"The final architecture (improved head, three WMFBs, positions {P2,P3,P5}, Haar basis) was selected by maximizing mAP on M3FD, and M3FD is also one of the four datasets used to compute the 4.5% headline gain. This means the M3FD row of Table 1 is partly a product of tuning on that benchmark. The authors should either fix all design choices before evaluation, select hyperparameters on a held-out validation split and then report test performance, or show that the same choices are also optimal on DroneVehicle, LLVIP, and FLIR-Aligned.","section":"Supplementary E.2; Tables 5-7"},{"comment":"The description of the WMFB placement is internally inconsistent. The architecture is introduced as containing 'three WaveMamba Fusion Blocks,' and the text says the first two backbone layers produce f^2 features that undergo DWT, but it then says DWT and WMFB are applied to the fused features at the 3rd and 5th layers. The supplementary ablation (Table C) uses positions {P2,P3,P5}. This inconsistency must be corrected because the exact fusion locations are essential for reproducing the method.","section":"3.2"},{"comment":"No error bars or multiple-seed results are reported. Several design comparisons differ by only 0.1-0.4 mAP (e.g., wavelet bases in Table E), which is within typical run-to-run variation for YOLO training; without repeated runs, the claim that the chosen configurations are optimal is not statistically supported. Reporting mean and standard deviation over at least three seeds would also help quantify the reliability of the headline differences.","section":"4.1; Supplementary Tables C-E"}],"minor_comments":[{"comment":"There is a typo in 'genetate outputs' that should read 'generate outputs.'","section":"3.1"},{"comment":"The sentence 'surpassing the fourth-place method by 2.6% and 3.6%' is confusing because WaveMamba is first in the table; the intended meaning is presumably 'the next-best method' or a specific ranked baseline.","section":"4.2, DroneVehicle"},{"comment":"The reference list contains duplicate entries: [67] and [68] are the same paper, and [75] and [76] are also the same paper; these should be merged.","section":"References"},{"comment":"In the HFE formula, if the RGB and IR coefficients have exactly equal absolute values, both masks are zero and the fused coefficient is set to zero; the authors should state whether this edge case occurs in practice or modify the mask to break ties.","section":"3.3.2, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the comparison protocol: the headline 4.5% claim is measured against published numbers without evidence that the baselines were re-run under the authors' pipeline, and the architecture was tuned on M3FD, which is also one of the four headline datasets. I would ask the authors to provide a controlled comparison for at least the closest competitors, report multiple-seed statistics, and clarify the WMFB placement inconsistency before the numerical claims can be accepted. There is no internal inconsistency in the derivations, and I do not see evidence of misconduct; the issues are empirical support and reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: WaveMamba is a competent engineering paper with a plausible frequency-domain story and large reported gains, but the headline 4.5-point average mAP lead over SOTA is not established because the comparison is against literature numbers and no code or error bars are provided. I'd send it to review, but the referee should push for a re-run of baselines under a shared protocol.\n\nWhat's actually new: the specific combination of Haar DWT for splitting RGB/IR features into low/high sub-bands, a Mamba-based low-frequency fusion block with channel swapping and gated attention, and an absolute-max high-frequency fusion, all feeding an IDWT-augmented YOLOv8 head. I can't point to a prior paper doing exactly this for RGB-IR detection. The paper is clearly written and the ablations are unusually complete for this area: they ablate the head, the SFM/DFM split, the fusion strategies, WMFB placement, number of blocks, and wavelet basis.\n\nThe soft spots are real. The central claim is measured against published results from other papers; Section 4.1 gives only WaveMamba's own training settings and never states that the baselines were re-run under those settings. Table 7 is internally suggestive: the authors' own average-fusion baseline gets 83.2 mAP50 / 55.1 mAP on M3FD, while the cited MMFN and MRD-YOLO numbers are 86.2/57.4 and 86.6/59.3. Whether that difference is protocol or implementation, the gaps in Tables 1-4 could shift by several points if everyone used the same pipeline. Also, the final architecture choices (WMFB placement, number, wavelet basis) were selected on M3FD, which is also a headline benchmark; that inflates the result modestly. No error bars, no code, so we can't tell.\n\nI'd also note the DFM module is only compared against removing it, not against a simpler cross-attention or non-gated Mamba fusion, so the 'gated attention' claim is a bit under-tested. Minor stuff: 'VEIDA' should be 'VEDAI' in the supplementary.\n\nBottom line: for someone working on multimodal detection, this is a useful reference and a reasonable baseline for future wavelet/Mamba fusion work. It deserves a serious referee: the engineering is solid, the experiments are broad, and the concerns are about verification, not about internal coherence. I'd make code release a condition of acceptance and ask the authors to re-run at least the closest baselines (MMFN, MRD-YOLO, ICAFusion) under their own training protocol, or to publish a detailed protocol comparison if re-running is infeasible.","headline":"WaveMamba is a competent engineering paper with a plausible frequency-domain story and large reported gains, but the headline 4.5-point mAP lead is not established because the comparison is against literature numbers and no code or error bars are provided.","tokens_in":21164,"tokens_out":3068,"would_cite":false,"duration_ms":32451,"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":"WaveMamba claims that fusing RGB and infrared imagery in the wavelet domain — using a Mamba-based block for low-frequency structure and an absolute-maximum selection for high-frequency detail — yields state-of-the-art object detection…","keywords":["RGB-infrared object detection","multi-spectral fusion","discrete wavelet transform","Mamba","state space model","feature fusion","YOLOv8","low-light detection"],"falsifier":"Re-run the compared methods (for example DMM, ICAFusion, CrossFormer, and the YOLOv8 single-modality baselines) in the same codebase, hyperparameters, and data splits used for WaveMamba on M3FD and DroneVehicle; if the best re-run baseline closes the gap to well below 4.5% or changes the ranking, the headline claim fails. A second check: measure low- and high-frequency information entropy on each new dataset — the motivational premise that IR dominates low frequency and RGB dominates high frequency is directly testable.","tokens_in":20131,"feed_emoji":"🎯","tokens_out":7422,"duration_ms":64564,"temperature":0.7,"pith_summary":"This paper argues that visible (RGB) and infrared (IR) images complement each other in the frequency domain, not just spatially: IR carries more information in the low-frequency sub-band, while RGB carries more in the high-frequency detail sub-bands. Building on that observation, WaveMamba decomposes both modalities with a discrete wavelet transform, fuses the low-frequency parts with a Mamba-based block that swaps channels and applies gated attention, and fuses the high-frequency parts by keeping whichever pixel has the larger absolute value. An inverse-wavelet detection head reconstructs the final features. On four public RGB-IR benchmarks (M3FD, DroneVehicle, LLVIP, FLIR-Aligned), the paper reports state-of-the-art mAP50 and mAP numbers, with an average gain of 4.5 points over the second-best method. If the comparisons hold up, this offers a general fusion recipe that works across three different backbones and improves detection in low light, bad weather, and dense small-target scenes.","feed_headline":"WaveMamba fusion lifts RGB-IR detection by 4.5% average mAP","feed_subtitle":"Infrared carries low-frequency structure; RGB carries detail — fusing each where it is strongest tops four benchmarks.","key_machinery":"The load-bearing mechanism is the WaveMamba Fusion Block (WMFB), which consists of the Low-frequency Mamba Fusion Block (LMFB) and the High-Frequency Enhancement (HFE) strategy. LMFB first performs a channel-swapping operation between the RGB and IR low-frequency sub-bands, passes each through Vision State Space (VSS) blocks, then runs a gated-attention deep fusion that alternates the modalities as primary and auxiliary and uses the Selective Scan 2D (SS2D) module — the core of Mamba — to propagate global information. HFE does no learned fusion at all: it builds a binary mask from the sign of the difference between the absolute values of the two modalities' high-frequency sub-images, keeping the larger-magnitude pixel from either RGB or IR and suppressing the other. Around these blocks, a discrete Haar wavelet transform splits features at multiple network depths into four sub-bands (LL, LH, HL, HH) and an inverse transform in the YOLOv8 head reconstructs the features for detection. Together they realize the argument that fusion should be frequency-aware — structure from IR, detail from RGB, selected rather than averaged.","core_discovery":"The central claim, on the paper's own terms, is that modality-specific frequency decomposition is the right organizing principle for RGB-IR fusion. The authors show by information-entropy measurement that IR dominates the LL sub-band and RGB dominates the LH/HL/HH sub-bands, then exploit this by routing low-frequency features through the WaveMamba Fusion Block and high-frequency features through an absolute-maximum selection mask. The Mamba-based low-frequency path (a shallow channel-swapping stage followed by a gated-attention selective-scan stage) is designed to use the global modeling strength of state space models on structure-rich low-frequency content, while the absolute-maximum rule preserves the sharpest edges and textures from either modality. Recombining via IDWT instead of standard upsampling is claimed to cut information loss and also shrink parameter count. The result is a claimed state-of-the-art performance across four benchmarks with an average mAP improvement of about 4.5 points over the previous best.","pith_inferences":["The frequency-entropy argument suggests the approach could extend to other modality pairs with complementary spectral content, such as RGB-depth or RGB-event streams, but the paper does not test this.","Because the reported gains are computed against published baseline numbers rather than re-run baselines, the 4.5% headline figure may shift under a unified training protocol; a fair head-to-head re-run is the natural next check.","The wavelet-basis ablation shows less than 0.4% sensitivity to the choice of wavelet, hinting that other multiresolution decompositions might behave similarly; that robustness is a testable conjecture, not a paper claim.","The IDWT head improves accuracy while cutting 7.6M parameters versus the standard YOLOv8 head, which suggests the lossless up/down-sampling benefit may apply to single-modality detectors as well."],"forward_implications":["On M3FD the YOLOv8-based WaveMamba reports 92.1 mAP50 and 64.4 mAP, gains of 5.5 and 5.1 points over the previous best, with the YOLOv5 and ResNet50 versions also leading their backbone groups.","On DroneVehicle the YOLOv8 variant reaches 79.8 mAP50 and 60.5 mAP with no remote-sensing-specific design, suggesting the frequency-fusion recipe transfers to dense small-target drone imagery.","On LLVIP pedestrian detection, the YOLOv8 model reports 98.3 mAP50 and 66.0 mAP, the best among compared fusion and single-modality methods.","On FLIR-Aligned, WaveMamba beats the compared methods in precision, recall, F1, mAP50, and mAP while using fewer parameters (e.g., 45.6M vs 340.0M for CrossFormer) and lower inference time.","The ablations attribute the gains to the design: removing SFM or DFM costs about 2 points of mAP, and the HFE+LMFB combination outperforms averaging fusion in the wavelet domain."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space model that the low-frequency fusion path builds on.","marker":"[15]"},{"why":"Defines the VSS block and 2D Selective Scan that the SFM and DFM modules use for visual feature sequences.","marker":"[40]"},{"why":"Motivates multi-level wavelet transforms for lossless sampling and large receptive fields, the structural precedent for WaveMamba's DWT placement.","marker":"[11]"},{"why":"Provides the channel-swapping operation used in the shallow fusion stage.","marker":"[17]"},{"why":"YOLOv8 is the base detector whose backbone, training pipeline, and original head WaveMamba modifies.","marker":"[60]"},{"why":"M3FD, the main benchmark dataset used in ablations and the SOTA comparison on extreme-weather multi-class detection.","marker":"[38]"},{"why":"DroneVehicle, the drone RGB-IR vehicle benchmark where the method is evaluated.","marker":"[53]"},{"why":"LLVIP, the low-light pedestrian benchmark where WaveMamba reports 98.3 mAP50.","marker":"[21]"},{"why":"FLIR-Aligned, the aligned FLIR benchmark with precision/recall/F1 evaluation.","marker":"[87]"}],"fun_headline_variants":["WaveMamba fuses RGB-IR by frequency, beats SOTA by 4.5% mAP","Wavelet-Mamba fusion: IR for low-freq, RGB for detail, +4.5% mAP","Frequency-first fusion: WaveMamba tops four RGB-IR benchmarks","Mamba meets wavelet: RGB-IR detection gains 4.5 mAP on 4 datasets","Split by frequency, fuse with Mamba: WaveMamba wins by 4.5% mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 4.5% average improvement is measured against numbers the paper takes from earlier publications, without re-running those methods under its own training, split, and augmentation settings.","fun_headline_variants_meta":{"raw":{"variants":["WaveMamba fuses RGB-IR by frequency, beats SOTA by 4.5% mAP","Wavelet-Mamba fusion: IR for low-freq, RGB for detail, +4.5% mAP","Frequency-first fusion: WaveMamba tops four RGB-IR benchmarks","Mamba meets wavelet: RGB-IR detection gains 4.5 mAP on 4 datasets","Split by frequency, fuse with Mamba: WaveMamba wins by 4.5% mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000957,"raw_usage":{"total_tokens":4079,"prompt_tokens":945,"completion_tokens":3134,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":3012}},"tokens_in":561,"tokens_out":3134,"duration_ms":21566,"temperature":1.0,"reasoning_tokens":3012,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:38:57.923798+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the compared methods (for example DMM, ICAFusion, CrossFormer, and the YOLOv8 single-modality baselines) in the same codebase, hyperparameters, and data splits used for WaveMamba on M3FD and DroneVehicle; if the best re-run baseline closes the gap to well below 4.5% or changes the ranking, the headline claim fails. A second check: measure low- and high-frequency information entropy on each new dataset — the motivational premise that IR dominates low frequency and RGB dominates high frequency is directly testable.","supporting_citations":[{"cited_title":"Yolov8: A novel object detection algorithm with enhanced performance and robust- ness","cited_arxiv_id":null,"evidence_quote":"YOLOv8 is the base detector whose backbone, training pipeline, and original head WaveMamba modifies."},{"cited_title":"Target- aware dual adversarial learning and a multi-scenario multi- modality benchmark to fuse infrared and visible for object detection","cited_arxiv_id":null,"evidence_quote":"M3FD, the main benchmark dataset used in ablations and the SOTA comparison on extreme-weather multi-class detection."},{"cited_title":"Drone-based rgb-infrared cross-modality vehicle detection via uncertainty-aware learning","cited_arxiv_id":null,"evidence_quote":"DroneVehicle, the drone RGB-IR vehicle benchmark where the method is evaluated."},{"cited_title":"Multispectral fusion for object detection with cyclic fuse-and-refine blocks","cited_arxiv_id":null,"evidence_quote":"FLIR-Aligned, the aligned FLIR benchmark with precision/recall/F1 evaluation."}],"review_version":1}