{"id":"30cc9324-1931-4145-90a3-477966f3d440","arxiv_id":"2507.00849","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"UAVD-Mamba, a Mamba-based multimodal detector with deformable token blocks, reaches 83.0% mAP on DroneVehicle, outperforming OAFA by 3.6%.","lead":"UAVD-Mamba is a neural network that detects vehicles and drones by combining visible and infrared images using a Mamba architecture. On the DroneVehicle benchmark it reports 83.0% mean average precision, beating the previous best method OAFA by 3.6 percentage points.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OAFA's DroneVehicle numbers may use a different ground-truth annotation rule than the paper's, so the 3.6% mAP advantage over OAFA is not verified as a fair comparison.","rationale":"The reader's weakest_assumption correctly identifies the protocol risk as the load-bearing issue. The paper's own admission that OAFA is not open-source (Fig. 4 caption) makes the imported 79.4 number unverifiable, and the paper does not cite OAFA among the protocol sources [24], [31]. An additional internal ambiguity—the ablation base of 79.6 not matching the YOLOv11 IR value of 79.3 in Table I—is secondary; even if resolved, it does not settle whether the external comparison to OAFA is fair. Thus the conditional verdict stands pending a protocol-equivalence check.","tokens_in":10378,"tokens_out":13081,"duration_ms":128132,"concrete_test":"Run UAVD-Mamba's trained model on the DroneVehicle test set and evaluate it under two ground-truth constructions: (i) the paper's 'modality with more annotations' rule and (ii) the union of RGB and IR annotations (a common alternative that OAFA may have used). If the mAP under rule (ii) is not at least 3.6 points above OAFA's reported 79.4, or if the two rules differ by more than 0.5 mAP, then determine OAFA's exact rule from its paper/code; if it differs from the paper's, re-run the comparison under that rule and check whether the 3.6% advantage is preserved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim (abstract and Section IV-B) is that UAVD-Mamba outperforms OAFA by 3.6% mAP (83.0 vs 79.4). Section IV-A specifies the evaluation protocol: ground truth is taken from the modality image with more annotations, and mAP@0.5 with OBB is reported, 'following previous studies [24], [31]'. OAFA [25] is not among those citations, and the paper notes (Fig. 4 caption) that OAFA is not open-source, so OAFA's 79.4 is imported from its paper without re-implementation. If OAFA used a different annotation merging rule (e.g., union of RGB and IR labels, or the official DroneVehicle benchmark rule), the comparison is not apples-to-apples. Because the paper's own base YOLOv11 (RGB+IR, if that is the ablation base) already reaches 79.6—slightly above OAFA—the entire 3.6% advantage could be an artifact of protocol differences rather than the method's merit. This concern is load-bearing because the central claim of SOTA performance rests entirely on this comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces UAVD-Mamba, a dual-stream RGB-infrared object detector for UAV imagery. The method combines deformable-token Vision Mamba blocks (DTMB), a Fusion Mamba Block for cross-modal state-space fusion, cross-enhanced spatial and cross-channel attention, a four-scale stack of DTMBs, and a YOLOv11-inspired detection neck (DNM). On the DroneVehicle dataset the paper reports 83.0% mAP@0.5, which is 3.6 points above the OAFA baseline, and an ablation study attributes the gains to DTMB, FFAR, and DNM. Source code is promised but not yet provided.","tokens_in":10605,"tokens_out":8388,"duration_ms":99631,"significance":"If the headline comparison is protocol-fair, the paper is a useful applied contribution: the deformable-token mechanism is a plausible remedy for Mamba's fixed tokenization, the architecture is clearly motivated, the efficiency numbers are informative, and the paper includes a candid limitation discussion. The evaluation is against an external benchmark, and I found no circularity in the core empirical claim. However, the central SOTA claim rests entirely on the comparison to OAFA, whose evaluation protocol is not verified to match the one used here, and the ablation gains are small and reported without variance information. The paper therefore needs a strengthening of the experimental evidence before its claims can be accepted.","major_comments":[{"comment":"The 3.6% mAP improvement over OAFA is not verified as a fair comparison. Section IV-A defines the evaluation protocol (ground truth taken from the modality with more annotations, mAP@0.5 with oriented bounding boxes) and cites [24] and [31] as precedents; OAFA [25] is not among those citations. The Fig. 4 caption states that OAFA is not open-source, so the 79.4% OAFA number is imported without re-implementation. If OAFA uses a different annotation-merging rule, evaluation split, or metric convention, the entire headline advantage could be an artifact; this concern is amplified by the fact that the paper's own YOLOv11 base reaches 79.6% in the ablation (Table IV), already above OAFA's 79.4%. Please confirm OAFA's exact protocol, re-run it under the same protocol, or obtain and evaluate its predictions with the same evaluation code, and report the corrected comparison.","section":"Section IV-A; Table I; Fig. 4"},{"comment":"The ablation does not cleanly isolate the DTMB contribution because FFAR is defined in Section III-B as containing the Cross-enhanced Spatial Attention, DTMB, Fusion Mamba Block, and Cross-Channel Attention. Thus the 'Base+D+F' configuration adds a second DTMB (the one inside FFAR) on top of the standalone DTMB, so the +2.7% gain cannot be attributed to FFAR alone, and the conclusion that DTMB contributes most is not supported by the reported decomposition. Please clarify whether the standalone DTMB in the ablation is the same component as the DTMB inside FFAR, or provide a configuration that removes one of the two DTMBs, and describe the exact module stack for each row. In addition, the ablation base row reports 79.6% while Table I lists YOLOv11 (Base) with RGB+IR at 79.3%, so the base configuration needs to be stated precisely.","section":"Section III-B; Table IV"},{"comment":"The empirical claims lack statistical support and full reproducibility. The paper reports a single run with no error bars or significance tests; the ablation margins are 0.6-0.7 mAP, which may be within run-to-run noise, especially for a 100-epoch training schedule. The loss weights lambda_cls, lambda_box, and lambda_dfl in Eq. (11) are never specified, and the stride, kernel size, and token-sequence construction inside the DTMB are not fully described. Please report mean and standard deviation over multiple seeds (or a significance test), provide the exact loss weights, and give the missing DTMB configuration details.","section":"Section IV-A, Section IV-E, Eq. (11)"}],"minor_comments":[{"comment":"Eq. (7) says 'AvgPool(\\cdot) denote maximum pooling', which is contradictory; please use either average pooling or maximum pooling consistently in the formula and the text.","section":"Section III-B, Eq. (7)"},{"comment":"The sentence claiming 'fewer parameters and GFlops' is inaccurate relative to SLBAF, which has 6.3M parameters compared to the proposed 39.7M; please qualify the claim to the compared methods that the paper actually outperforms in both accuracy and efficiency.","section":"Section IV-C, Table II"},{"comment":"There is a typo in Eq. (11): 'lambda_clc' should be 'lambda_cls' for the classification loss coefficient.","section":"Section III-D, Eq. (11)"},{"comment":"The citation mapping does not match the bibliography: the text attributes a Cross-modal Fusion Mamba (CFM) module to [11], but [11] in the reference list is 'Coupled Mamba', while [12] is listed as 'Remotedet-Mamba' even though the text says [12] applied Coupled Mamba; please correct the citation labels.","section":"Section II-B"},{"comment":"The formatting of Table I is difficult to follow for the single-modality rows: DAIK and I2MDet appear under the RGB input column but their Input field says IR; please align the table so each method's input modality is unambiguous.","section":"Table I"},{"comment":"The description of how normal patches and adaptive patches are merged into one token sequence is informal; Eq. (3) states that the two convolution outputs are added, but the kernel sizes, strides, and how the resulting tensor is flattened into Mamba tokens should be specified for reproducibility.","section":"Section III-B and Section IV-D"}],"recommendation":"major_revision","confidential_remarks":"This is a competent applied paper, and the central architecture is defensible, but the headline SOTA claim depends on a protocol match with OAFA that is not demonstrated. I would urge the editor to require the authors to either re-run OAFA under their exact protocol or provide a verified protocol match, and to add multi-seed results. If these issues are resolved, the paper could be acceptable; as it stands, the empirical foundation is too fragile for acceptance. The paper fits the scope of an applied computer-vision venue, and the promised code release would help substantially."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the deformable token Mamba block (DTMB) is a real architectural novelty: feeding adaptive patches from deformable convolutions plus normal patches into a Mamba block is not in the cited prior work, and the ablation shows it buys +2.1 mAP over the YOLOv11 base. Second, the headline claim of 3.6% over OAFA is not verified: OAFA is not open-source, its number is imported from its paper, and the paper doesn't confirm OAFA used the same ground-truth annotation rule (labels from the modality with more annotations) and mAP@0.5 protocol. Since the paper's own base (YOLOv11 RGB+IR) already reaches 79.6, slightly above OAFA's 79.4, all of the claimed advantage could be protocol artifact.\n\nThe paper does well on structure: equations for the cross-attention and fusion blocks are clear, the multiscale stacking and neck modifications are concrete, and the limitation section honestly flags the freight-car weakness. That is more than many detection papers do.\n\nSoft spots, in order. The OAFA comparison is load-bearing and unconfirmed; fix requires either re-running OAFA under the same protocol or stating explicitly that the annotation rule and metric match. The paper lacks error bars or statistical tests; one run per configuration, so the +2.1% DTMB gain could be noise. Loss coefficients λ_cls, λ_box, λ_dfl are defined but values never given. The parameter-efficiency sentence overstates: SLBAF has 6.3M params versus 39.7M, so 'fewer parameters' only holds against the bigger Transformer models, not SLBAF. Minor: the visualization of adaptive patches is illustrative, not quantitative.\n\nWho is it for: people working on Mamba-based detection or RGB-IR fusion for aerial vehicles. The DTMB idea is worth citing even if the SOTA number collapses. It deserves a serious referee, because the architecture is new and the internal evidence is coherent; but the paper needs code, multiple seeds, protocol confirmation, and the parameter comparison corrected before acceptance.","headline":"A genuinely new deformable-token Mamba block that appears to help, but the headline 3.6% edge over OAFA is not verified because the evaluation protocol comparison is unconfirmed.","tokens_in":11126,"tokens_out":2199,"would_cite":true,"duration_ms":24309,"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":"UAVD-Mamba, a Mamba-based RGB-infrared detector with deformable tokens, reports 83.0% mAP on DroneVehicle, 3.6 points above the OAFA baseline.","keywords":["multimodal UAV detection","Mamba","deformable token","RGB-infrared fusion","oriented object detection","DroneVehicle dataset","feature fusion","state space model"],"falsifier":"Run UAVD-Mamba and OAFA on the same DroneVehicle test split with the same ground-truth merging rule and the same mAP@0.5 orientation-aware evaluation script. If UAVD-Mamba's mAP does not come out 3.6 points above OAFA's re-measured mAP, the headline comparison fails; a second check is to retrain the base-plus-DTMB configuration without the deformable-convolution branch and see whether mAP drops by roughly the 2.1 points reported in the ablation.","tokens_in":1953,"feed_emoji":"🚁","tokens_out":2279,"duration_ms":107631,"temperature":0.7,"pith_summary":"The paper sets out to show that Mamba-style state-space models can handle the irregular, occluded, small targets typical of UAV imagery if the tokens fed into the sequence model are allowed to deform rather than staying on a fixed square grid. It builds UAVD-Mamba, a dual-stream RGB and infrared detector, where each modality gets its own Deformable Token Mamba Block, the two streams meet in a Fusion Mamba Block, and multiscale features are processed by a YOLOv11-inspired neck adapted for Mamba. On the DroneVehicle dataset the method reports 83.0% mAP, 3.6 points higher than the OAFA baseline, and the ablation attributes the largest single gain to the deformable token block. If the result holds, it suggests that Mamba's efficient long-range modeling can be made geometrically adaptive enough for practical aerial detection.","feed_headline":"Deformable Mamba tokens push UAV detection to 83.0% mAP","feed_subtitle":"RGB and infrared fusion through Mamba tokens surpasses the prior multimodal best on DroneVehicle by 3.6 points.","key_machinery":"The central object is the Deformable Token Mamba Block (DTMB), which generates deformable tokens by adding the outputs of a standard convolution and a deformable convolution, $T_m = \\operatorname{Conv}(F^{cs}_m) + \\operatorname{DConv}(F^{cs}_m)$, and feeds these tokens into a Vision Mamba Block. This gives the state-space model patches that can adapt their geometry to irregular objects rather than fixed square cells. Around this, the Fusion Mamba Block extends the state-space model to a two-input fusion form so that RGB and IR features supplement each other, and the cross-channel attention $F_f = F^{FM}_{rgb} \\times F^c_{rgb}/F^c_{ir} + F^{FM}_{ir} \\times F^c_{ir}/F^c_{rgb}$ is the mechanism that removes cross-modal redundancy. A multiscale variant stacks four DTMBs, feeding three of their outputs into the Detection Neck for Mamba, which replaces YOLOv11's C3K2 module with Mamba blocks and applies SPPF followed by Mamba at each scale.","core_discovery":"The paper's central claim is that a Mamba-based multimodal detector constructed around deformable tokens reaches 83.0% mAP on DroneVehicle, outperforming the OAFA baseline at 79.4% by 3.6 points, while also improving on per-class accuracy for cars, trucks, buses, and vans. The claim is carried by the Deformable Token Mamba Block, which forms each input token by adding a normal patch from a standard convolution to an adaptive patch from a deformable convolution, so that the token shape follows image content instead of a fixed division. Two separate DTMBs process the RGB and infrared streams, a Fusion Mamba Block exchanges state between the modalities through a two-input state-space model, and a cross-channel attention formula divides each modality's feature by the other modality's channel attention to suppress redundancy. The paper further reports that adding DTMB to a YOLOv11 base raises mAP from 79.6% to 81.7%, that the full pipeline reaches 83.0%, and that a faster variant with only DTMB still reaches 81.7% mAP at higher frame rates.","pith_inferences":["A direct head-to-head that re-runs OAFA with the same DroneVehicle split, ground-truth merging rule, and mAP@0.5 evaluation script would confirm whether the 3.6-point gap survives protocol matching; the paper follows a prior protocol but does not verify that OAFA used it.","The deformable-token construction could be ported to other Mamba-based vision backbones and to tasks with irregular objects, such as medical image segmentation or remote-sensing instance detection, where fixed patch grids are known to lose information.","The cross-channel attention formula, which divides by the other modality's channel attention, is a general anti-redundancy operation that could be tested on other two-stream fusion problems, especially with noisy or misaligned modality pairs."],"forward_implications":["On the DroneVehicle dataset, UAVD-Mamba reaches 83.0% mAP, 3.6 points above the OAFA baseline, with the highest per-class scores among compared detectors on car, truck, bus, and van.","The deformable token construction is the main source of gain: adding DTMB to a YOLOv11 base raises mAP from 79.6% to 81.7%, a larger jump than the later additions of FFAR and DNM.","Each proposed module contributes: FFAR adds another 0.7 points on top of DTMB, and DNM adds 0.6 points on top of DTMB plus FFAR, bringing the final total to 83.0%.","A faster configuration, UAVD-Mamba-FAST using only DTMB, retains 81.7% mAP while running at 45.0 FPS on an A6000 and 24.2 FPS on an RTX 4090, indicating a practical speed-accuracy tradeoff.","The full model uses 39.7 million parameters and 38.9 GFLOPs, which is lighter than several compared multimodal baselines while achieving higher mAP."],"supporting_citations":[{"why":"Supplies the OAFA baseline whose 79.4% mAP is the direct comparison target for the headline 3.6-point improvement.","marker":"[25]"},{"why":"Provides the DroneVehicle dataset with 28,439 RGB-infrared image pairs used for training, validation, and testing.","marker":"[32]"},{"why":"Supplies the Mamba state-space model that the entire detector framework is built on.","marker":"[10]"},{"why":"Supplies the deformable convolution used inside DTMB to generate adaptive patches.","marker":"[28]"},{"why":"Provides the Vision Mamba Block implementation that DTMB feeds into and that the Fusion Mamba Block extends.","marker":"[29]"},{"why":"Defines the YOLOv11 base detector, loss function, and the C3K2, SPPF, and detection head components that are modified into the Detection Neck for Mamba.","marker":"[30]"},{"why":"One of the prior studies whose evaluation protocol for mAP@0.5 and ground-truth merging rule the paper follows.","marker":"[24]"},{"why":"Another prior study cited alongside [24] for the evaluation protocol and ground-truth selection rule.","marker":"[31]"}],"fun_headline_variants":["Deformable tokens let Mamba fuse RGB-IR for 83.0% mAP","RGB-IR Mamba with deformable tokens hits 83.0% mAP","Deformable Mamba tokens improve UAV detection by 3.6% mAP","Mamba token fusion beats prior multimodal UAV detector by 3.6%"],"cache_read_input_tokens":13312,"weakest_assumption_plain":"The 3.6% improvement over OAFA is valid only if both methods were measured on the same test images with the same scoring rule; the paper follows a protocol used by earlier studies but does not confirm that OAFA used exactly that protocol.","fun_headline_variants_meta":{"raw":{"variants":["Deformable tokens let Mamba fuse RGB-IR for 83.0% mAP","RGB-IR Mamba with deformable tokens hits 83.0% mAP","Deformable Mamba tokens improve UAV detection by 3.6% mAP","Mamba token fusion beats prior multimodal UAV detector by 3.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00065,"raw_usage":{"total_tokens":3076,"prompt_tokens":1131,"completion_tokens":1945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":747,"completion_tokens_details":{"reasoning_tokens":1856}},"tokens_in":747,"tokens_out":1945,"duration_ms":16459,"temperature":1.0,"reasoning_tokens":1856,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:04:57.735058+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run UAVD-Mamba and OAFA on the same DroneVehicle test split with the same ground-truth merging rule and the same mAP@0.5 orientation-aware evaluation script. If UAVD-Mamba's mAP does not come out 3.6 points above OAFA's re-measured mAP, the headline comparison fails; a second check is to retrain the base-plus-DTMB configuration without the deformable-convolution branch and see whether mAP drops by roughly the 2.1 points reported in the ablation.","supporting_citations":[{"cited_title":"Weakly misalignment-free adaptive feature alignment for uavs-based multimodal object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the OAFA baseline whose 79.4% mAP is the direct comparison target for the headline 3.6-point improvement."},{"cited_title":"Drone-based rgb-infrared cross-modality vehicle detection via uncertainty-aware learning,","cited_arxiv_id":null,"evidence_quote":"Provides the DroneVehicle dataset with 28,439 RGB-infrared image pairs used for training, validation, and testing."},{"cited_title":"Deformable convnets v2: More deformable, better results,","cited_arxiv_id":null,"evidence_quote":"Supplies the deformable convolution used inside DTMB to generate adaptive patches."},{"cited_title":"Fusionmamba: Efficient remote sensing image fusion with state space model,","cited_arxiv_id":null,"evidence_quote":"Provides the Vision Mamba Block implementation that DTMB feeds into and that the Fusion Mamba Block extends."},{"cited_title":"ultralytics/yolov11,","cited_arxiv_id":null,"evidence_quote":"Defines the YOLOv11 base detector, loss function, and the C3K2, SPPF, and detection head components that are modified into the Detection Neck for Mamba."},{"cited_title":"Multi-modal object detection of uav remote sensing based on joint representation optimization and specific information enhance- ment,","cited_arxiv_id":null,"evidence_quote":"One of the prior studies whose evaluation protocol for mAP@0.5 and ground-truth merging rule the paper follows."},{"cited_title":"Directional alignment instance knowledge distillation for arbitrary-oriented object detection,","cited_arxiv_id":null,"evidence_quote":"Another prior study cited alongside [24] for the evaluation protocol and ground-truth selection rule."}],"review_version":1}