{"id":"a1a5f49b-50a1-4e70-8597-1a7d625989b5","arxiv_id":"2506.16737","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CoDAF aligns and fuses visible and infrared UAV images with offset-guided deformable convolution and attention gating, reaching 78.6% mAP on DroneVehicle.","lead":"UAV object detection that combines visible and infrared images often fails when the two camera views are slightly misaligned. This paper proposes CoDAF, a network that aligns the image streams with deformable convolutions and blends them with adaptive gating, reporting top results on two benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The M3FD result (90.8/61.2, Table II) is produced on a custom 3,368/831 split rather than the protocol used by the quoted baselines, so the +1.0/+0.5 mAP margins over CFT are protocol-dependent and the M3FD prong of the SOTA claim is unverified.","rationale":"The paper's central claim is that CoDAF, by jointly aligning and fusing RGB/IR features, achieves state-of-the-art weakly aligned UAV detection on DroneVehicle and M3FD. For this claim to hold, the reported margins must be measured under protocols equivalent to those of the quoted baselines, and the experimental numbers must be reproducible from the manuscript. The DroneVehicle headline (78.6% mAP@.5, +4.0 over CALNet/CAGTDet) is the most defensible part of the paper: it uses the official benchmark protocol, and the ablation (Table III) isolates a +4.7 point gain of the proposed modules over the paper's own baseline. The IR-only annotation choice is common practice on this benchmark and is explicitly disclosed. Residual risks here are single-run variance and the fact that baselines are quoted rather than re-run, but neither is decisive given the 4.0-point margin. The M3FD prong is the weak link. Section IV-A2 discloses a custom 3,368/831 split built to maximize scene dissimilarity, and Table II compares CoDAF's numbers with published results obtained under the standard M3FD protocol. The margins over the strongest quoted baseline (CFT) are only +1.0 mAP@.5 and +0.5 mAP@.5:.95, an amount a split change can plausibly move. The exact image indices are not released, so no third party can reproduce or audit the comparison. If the margins vanish under a shared protocol, that prong of the claim fails. The manuscript also contains a direct internal contradiction that blocks replication of the main experiments: Eq. (19) sets lambda to 0.1, while the ablation in Section IV-E2 reports the best mAP at lambda = 0.3. This is a concrete inconsistency in the experimental specification that the authors must resolve, even though its resolution would likely not refute the DroneVehicle result. The reader's designated weakest assumption (Eq. (8), the SSIM+L1 alignment loss) is plausible as a design risk, but I judge it less load-bearing: the detection loss with a small lambda dominates training, so over-constraint that suppresses modality-specific cues would be partly counteracted by Ldet; and the paper's own (internally inconsistent) lambda sweep claims a moderate alignment weight is beneficial, implying the loss helps on net. Even if the concern landed, it would suggest suboptimality, not invalidate the SOTA result. I therefore keep the reader's CONDITIONAL verdict: the DroneVehicle result is reasonably supported, the M3FD comparison must be re-run on a shared split, the lambda discrepancy must be corrected, and code/split definitions should be released. The proposed 2x2 M3FD evaluation is the single settlement check.","tokens_in":19999,"tokens_out":27475,"duration_ms":268859,"concrete_test":"Publish the exact M3FD split (image indices) and run a two-by-two evaluation: (1) the public CFT code on CoDAF's 3,368/831 split, and (2) CoDAF on the official M3FD split; compare mAP@.5 and mAP@.5:.95 across the four runs. If the +1.0 / +0.5 point margins do not persist when both methods share a split, the M3FD prong of the SOTA claim fails. Repeat CoDAF with at least three seeds to bound run-to-run variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Two benchmarks back the SOTA claim, and they are not equally secure. The DroneVehicle headline (78.6%, Table I) uses the official test protocol and the ablation (+4.7 over the paper's own baseline) supports the modules, so I do not see a decisive threat there. The M3FD prong is the weak link: Section IV-A2 discloses a custom 3,368/831 split built on scene dissimilarity, and Table II compares against published numbers (CFT, E2E-MFD) obtained under the standard M3FD protocol. The claimed margins are only +1.0 mAP@.5 (90.8 vs 89.8) and +0.5 mAP@.5:.95 (61.2 vs 60.7) over the strongest quoted baseline, CFT, which is within the range that a change of evaluation split can shift detection mAP. The exact image indices are not released, so the comparison cannot be reproduced or audited. The problem is compounded by an internal inconsistency in the experimental specification: Eq. (19) states lambda = 0.1, while the ablation in Section IV-E2 reports the best mAP at lambda = 0.3, so the main experiments cannot be replicated as written. If the M3FD margin does not survive a shared-protocol comparison, that prong of the claim fails even if the DroneVehicle result stands.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses weakly aligned RGB-IR object detection for UAVs. It proposes CoDAF, which couples an Offset-guided Semantic Alignment (OSA) module with a Dynamic Attention-guided Fusion Module (DAFM) inserted at three backbone stages. OSA estimates attention-guided spatial offsets and uses DCNv2, with a shared semantic space (SID) and a contrastive loss, to align visible features to infrared features; DAFM combines modality-adaptive gating with spatial-channel cross-modulation for fusion. The total loss combines the detection loss with alignment losses. Experiments on DroneVehicle report 78.6% mAP@.5, which is 4.0 points above the best quoted competitor, and experiments on M3FD report 90.8% mAP@.5 and 61.2% mAP@.5:.95. Ablations on DroneVehicle attribute the gains to OSA, DAFM, the IR-based attention map, the contrastive loss, and the stage-wise placement of the modules.","tokens_in":20320,"tokens_out":10069,"duration_ms":100987,"significance":"If the empirical results hold, CoDAF is a strong candidate for weakly aligned multimodal UAV detection: the DroneVehicle margin is substantial, the ablation suite is fairly complete (component-wise, stage-wise, attention alternatives, and efficiency), and the method is clearly described. The paper also discloses its main limitations, including the IR-only label choice, the custom M3FD split, and computational cost. However, the M3FD results are currently not comparable to the quoted baselines because the evaluation protocol differs in both split and detector, and the main-loss hyperparameter in Eq. (19) contradicts the ablation in Section IV-E2. These issues are concentrated on the secondary benchmark, so the DroneVehicle claim remains plausible, but the full state-of-the-art claim cannot be accepted as written.","major_comments":[{"comment":"Table II is not a controlled comparison. Section III-D states that on M3FD the method uses the RT-DETR head and loss, while the table caption describes all competitors as image-fusion methods plus YOLOv5s; the quoted CFT and E2E-MFD numbers were published under that YOLOv5s protocol and, per Section IV-A2, under a different data split, namely the authors' custom 3,368/831 scene-dissimilar split, whose exact image indices are not released. The reported margins of +1.0 mAP@.5 and +0.5 mAP@.5:.95 over CFT are therefore within the range that protocol changes can shift detection mAP, and the M3FD prong of the state-of-the-art claim is unsupported as presented. Please retrain all competitors on the same split with the same detector, release the split indices, and also report the comparison under the standard M3FD split.","section":"§III-D, §IV-A2, §IV-D, Table II"},{"comment":"The manuscript specifies the balancing hyperparameter as empirically set to 0.1 in the total loss in Eq. (19), but the ablation in Section IV-E2 reports that the best mAP is obtained at lambda = 0.3. As written, the main experiments cannot be reproduced, and the headline numbers appear to use a value that the paper's own ablation says is suboptimal. Please state the exact lambda used for every experiment, reconcile the main configuration with the ablation, and report results for both 0.1 and 0.3.","section":"§IV-E2 and Eq. (19)"},{"comment":"The DroneVehicle training protocol uses only the IR ground-truth annotations, justified as 'more comprehensive and accurate.' The paper does not state whether the comparison methods were retrained with the same label source or whether the official test evaluation uses IR labels, visible labels, or both. Because OSA explicitly aligns visible features to the IR modality and the DAFM gates on modality contributions, training and evaluating with IR-only labels can favor the proposed design over methods trained under the standard protocol. Please clarify the label protocol for all compared methods and, if possible, report results with the alternative label choice.","section":"§IV-B"},{"comment":"All reported results are single runs with no variance or significance information. Given the small +1.0/+0.5 margins on M3FD and the non-standard M3FD protocol, it is difficult to judge whether the difference over CFT is meaningful. Please report at least three seeds with mean and standard deviation for the main experiments, or release predictions so that the exact comparisons can be reproduced.","section":"Tables I and II"}],"minor_comments":[{"comment":"The headings 'Anaysis of the Alignment Loss' and 'Anaysis of the DAFM' contain typos; 'Anaysis' should be 'Analysis.' In Section IV-E3, the text refers to 'MAGM,' which should be 'MAGN' to match the module name defined in Section III-C.","section":"§IV-E2 and §IV-E3"},{"comment":"Eq. (9) is described as a sparsity loss, but it is, up to constants, the entropy of the attention map; minimizing it drives entries toward 0 or 1 rather than directly enforcing spatial compactness. Either correct the formulation to a true sparsity penalty or soften the interpretation in the text.","section":"Eq. (9)"},{"comment":"Table VIII reports FPS without specifying the GPU model, batch size, or whether preprocessing and postprocessing are included; comparisons with published FPS numbers are therefore not meaningful as stated.","section":"§IV-E6, Table VIII"},{"comment":"The component-wise OSA ablation removes the contrastive loss, but it does not ablate the SSIM/L1 components of Lsm or the attention loss Lattn. An ablation of these terms would clarify whether the pixel-level matching in Eq. (8) helps or hurts detection.","section":"§IV-E1 and Eq. (8)"},{"comment":"The conclusion's limitation statement that the method has high computational cost sits somewhat uneasily with the real-time framing in the abstract and the FPS claims in Section IV-E6; please align the wording.","section":"Conclusion and Abstract"},{"comment":"There are several typos and spacing artifacts, including 'cross-moda' in the contribution statement, 'UA V' in the abstract and body, and 'compariosn' in Section IV-D; a copy-edit pass is needed.","section":"Throughout"},{"comment":"The paper says M3FD has no official split, but several published methods report results under a standard M3FD protocol. If such a split exists, the custom split should be aligned with it, or at minimum the authors should justify why the custom split is preferable and release the exact image indices for reproducibility.","section":"§IV-A2"}],"recommendation":"major_revision","confidential_remarks":"The main reason for the major revision is the M3FD protocol mismatch: the comparison uses a custom split and a different detector head relative to the quoted baselines, and the exact split indices are not released. If the authors cannot provide a shared-protocol comparison, the M3FD state-of-the-art claim should be withdrawn or explicitly relabeled as a non-standard-protocol result. The DroneVehicle contribution is sufficiently strong and internally consistent to merit revision rather than rejection. I would also encourage the editors to ask for the split indices and a statement on code release, since the custom M3FD split is otherwise unverifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CoDAF is a legitimately new integration of existing techniques—attention-derived offsets, DCNv2 with a base-offset prior, a shared contrastive space, and gated dual-attention fusion—applied to weakly aligned UAV RGB-IR detection. On DroneVehicle, the headline result (78.6% mAP@.5, +4.0 over prior best) survives scrutiny: official test protocol, complete ablations including a +4.7 gain over their own baseline, and a sanity check that removing modules at stage 1 hurts most. The M3FD prong is weaker. The paper discloses a custom 3,368/831 split based on scene dissimilarity, but then compares to published numbers from the standard M3FD protocol. The margins over CFT are only +1.0/+0.5 mAP, which a split change can easily absorb. That claim needs a same-protocol comparison, not a footnote.\n\nThere are two concrete fixable problems. First, Eq. (19) states lambda=0.1, while the ablation in Section IV-E2 reports the best mAP at lambda=0.3. As written, the main experiments cannot be reproduced with the stated setting. Second, all results are single-run, with no variance or significance. That is common in this area, but with a 4-point gain it matters less; with the 0.5-point M3FD margin it matters a lot.\n\nI also note two smaller things. Using IR-only annotations on DroneVehicle is disclosed but may not match how prior baselines were trained and evaluated. And the attention loss in Eq. (9) is effectively a negative entropy term, not obviously a sparsity prior as described. These are minor if the rest gets fixed.\n\nOverall, the core method is sound and the DroneVehicle result looks real. The paper needs revision, not rejection. A serious referee should ask for code, exact split definitions, the M3FD comparison under the standard protocol, and the lambda correction. I'd bring it to a reading group only after those are addressed.","headline":"CoDAF is a credible new alignment-fusion combination that earns its DroneVehicle SOTA claim, but the M3FD claim is undermined by a non-standard split and a lambda inconsistency blocks reproducibility.","tokens_in":20875,"tokens_out":2651,"would_cite":true,"duration_ms":25345,"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 jointly aligning RGB features to an infrared reference and then adaptively fusing them resolves the two failure modes of weak cross-modal alignment and achieves 78.6% mAP@.5 on DroneVehicle.","keywords":["multimodal object detection","RGB-infrared fusion","weak alignment","deformable convolution","attention mechanism","UAV object detection","contrastive learning","oriented object detection"],"falsifier":"Train CoDAF on DroneVehicle with the pixel-value term $L_{mae}$ removed from Eq. (8) (so only SSIM, contrastive, and attention losses remain) and compare mAP@.5 against the reported 78.6%; if the variant matches or exceeds it, the pixel-level similarity constraint is not doing the claimed work, and the paper's explanation of why the full loss helps would need revision.","tokens_in":19782,"feed_emoji":"🚁","tokens_out":9741,"duration_ms":82570,"temperature":0.7,"pith_summary":"Unmanned aerial vehicles that detect objects from both visible and infrared images face a problem the paper calls weak alignment: the two cameras rarely show the same object at the same pixel because the platform moves and the sensors capture at different moments. The paper claims that this misalignment produces two distinct failures—semantic inconsistency at corresponding locations and modality conflict during fusion—and that previous methods fix only one of them. It proposes CoDAF, a single framework that first aligns visible features onto the infrared reference using attention-guided offsets and deformable convolution, then fuses the two modalities with per-pixel gating and dual attention. The reported result is 78.6% mAP@.5 on DroneVehicle, 4.0 points above the best prior method, and 90.8% mAP@.5 / 61.2% mAP@.5:.95 on M3FD.","feed_headline":"78.6% mAP: one framework aligns then fuses drone RGB-IR views","feed_subtitle":"Correcting spatial offsets before fusion beats prior RGB-IR detectors by 4.0 mAP on DroneVehicle.","key_machinery":"The load-bearing object is the offset-guided alignment loop inside OSA: an infrared-derived spatial attention map $M$ gates the concatenated RGB–IR features, a two-layer convolutional network predicts a base offset $\\phi_c$, and deformable convolution v2 (a convolution whose sampling locations are learned per position) uses $\\phi_c$ plus a learned residual offset to warp the visible feature map onto the infrared grid. Around that loop, a Shared Information Extractor (SID) with an InfoNCE contrastive loss projects both modalities into one semantic space to stabilize offset prediction, and the fusion module DAFM gates per-pixel modality contributions and then applies channel and spatial attention to the fused map. The total objective combines the detection loss with a weighted alignment loss ($\\lambda = 0.1$) that includes the contrastive term, the SSIM/L1 spatial alignment term of Eq. (8), and attention sparsity and smoothness terms.","core_discovery":"The central claim is that for RGB–IR UAV detection, weak cross-modal alignment—not the diversity of the modalities themselves—is the limiting factor, and its two symptoms, semantic inconsistency and modality conflict, must be treated jointly rather than in isolation. CoDAF implements this by making the infrared feature map the alignment reference: an attention map derived from the infrared features gates the concatenated RGB–IR features, a compact convolutional network predicts per-pixel offsets from that gated representation, and deformable convolution v2 uses those offsets as a base prior to resample the visible features onto the infrared grid. A shared feature extractor with an InfoNCE contrastive loss then pulls the aligned visible and infrared features into a common semantic space, while the fusion side assigns per-pixel weights to each modality and refines the fused map with channel-then-spatial attention. The paper reports that this joint design reaches 78.6% mAP@.5 on DroneVehicle, 4.0 points above the best prior method, and 90.8% mAP@.5 with 61.2% mAP@.5:.95 on M3FD.","pith_inferences":["If Eq. (8)'s pixel-level matching is over-constrained, a variant that drops the L1 term or uses a perceptual or feature-level similarity could preserve modality-specific cues; the paper does not ablate the individual coefficients $\\lambda_1$ and $\\lambda_2$ to test this.","Because the contrastive loss uses global pooled features, a patch-level or region-level contrastive objective might specifically help small and distant objects, the category where CoDAF falls slightly behind CALNet on freight-cars.","The 'align with the more reliable modality, then fuse adaptively' recipe is general: SAR–optical or RGB–depth pairs with registration errors are natural testbeds, with the modality that is geometrically stable serving as the reference."],"forward_implications":["Solving alignment and fusion in one network, with infrared as the alignment reference, raises weakly aligned RGB–IR UAV detection to 78.6% mAP@.5 on DroneVehicle, 4.0 points above the best prior method.","Replacing the infrared-derived attention map with a visible-derived one costs 1.5 mAP points in the paper's ablation, indicating that infrared is the more reliable spatial reference in low-light and cluttered scenes.","Removing OSA and DAFM at the first backbone stage drops accuracy from 78.6% to 75.4% mAP, more than at later stages, so early-stage alignment and fusion carry most of the benefit.","The same framework transfers to M3FD with a different detection head, reaching 90.8% mAP@.5 and 61.2% mAP@.5:.95, so the design is not tied to one detector or dataset.","CoDAF keeps parameters low (67.3M) and inference fast (58.1 FPS) despite higher FLOPs, indicating the alignment–fusion modules add modest computational cost for the accuracy gain."],"supporting_citations":[{"why":"Supplies the base-offset prior idea that CoDAF uses to regularize deformable-convolution alignment.","marker":"[55]"},{"why":"Defines deformable convolution v2, the operator whose sampling locations the predicted offsets modulate.","marker":"[56]"},{"why":"Provides the InfoNCE objective used to pull paired visible and infrared features into a shared semantic space.","marker":"[57]"},{"why":"Defines SSIM, the structural component of the spatial alignment loss in Eq. (8).","marker":"[58]"},{"why":"Provides the RT-DETR neck and detection head used on M3FD, grounding the detector framework and FPS claims.","marker":"[30]"},{"why":"Introduces DroneVehicle, the main benchmark where CoDAF reports 78.6% mAP@.5; also documents the misalignment statistics.","marker":"[9]"},{"why":"Introduces M3FD, the second benchmark where CoDAF reports 90.8% mAP@.5 and 61.2% mAP@.5:.95.","marker":"[17]"},{"why":"CALNet is the best prior method on DroneVehicle (74.6% mAP), the comparison behind the claimed 4.0-point gain.","marker":"[12]"},{"why":"CFT is the strongest prior fusion method on M3FD (60.7% mAP@.5:.95) that CoDAF surpasses.","marker":"[26]"}],"fun_headline_variants":["CoDAF: joint alignment + fusion lifts drone RGB-IR mAP to 78.6%","Weak alignment fixed: CoDAF aligns then fuses RGB-IR, hits 78.6% mAP","Unified RGB-IR alignment and fusion: CoDAF beats prior by 4.0 mAP","CoDAF jointly handles misalignment and fusion for drone detection: 78.6% mAP","Align then fuse: CoDAF's joint design boosts RGB-IR drone detection to 78.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the claim that, once geometrically corrected, visible features should closely match infrared features in both structure and pixel values (the $L_{sm}$ term of Eq. (8)); because RGB and infrared record different physical signals, this constraint could be too strong and suppress modality-specific details that detection needs.","fun_headline_variants_meta":{"raw":{"variants":["CoDAF: joint alignment + fusion lifts drone RGB-IR mAP to 78.6%","Weak alignment fixed: CoDAF aligns then fuses RGB-IR, hits 78.6% mAP","Unified RGB-IR alignment and fusion: CoDAF beats prior by 4.0 mAP","CoDAF jointly handles misalignment and fusion for drone detection: 78.6% mAP","Align then fuse: CoDAF's joint design boosts RGB-IR drone detection to 78.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1622,"prompt_tokens":1019,"completion_tokens":603,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":474}},"tokens_in":635,"tokens_out":603,"duration_ms":6251,"temperature":1.0,"reasoning_tokens":474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:19:29.582412+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CoDAF on DroneVehicle with the pixel-value term $L_{mae}$ removed from Eq. (8) (so only SSIM, contrastive, and attention losses remain) and compare mAP@.5 against the reported 78.6%; if the variant matches or exceeds it, the pixel-level similarity constraint is not doing the claimed work, and the paper's explanation of why the full loss helps would need revision.","supporting_citations":[{"cited_title":"Weakly misalignment-free adaptive feature alignment for uavs-based multimodal object detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the base-offset prior idea that CoDAF uses to regularize deformable-convolution alignment."},{"cited_title":"Deformable convnets v2: More deformable, better results,","cited_arxiv_id":null,"evidence_quote":"Defines deformable convolution v2, the operator whose sampling locations the predicted offsets modulate."},{"cited_title":"Detrs beat yolos on real-time object detection,","cited_arxiv_id":null,"evidence_quote":"Provides the RT-DETR neck and detection head used on M3FD, grounding the detector framework and FPS claims."},{"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":"Introduces M3FD, the second benchmark where CoDAF reports 90.8% mAP@.5 and 61.2% mAP@.5:.95."},{"cited_title":"Multispectral object detection via cross-modal conflict-aware learning,","cited_arxiv_id":null,"evidence_quote":"CALNet is the best prior method on DroneVehicle (74.6% mAP), the comparison behind the claimed 4.0-point gain."}],"review_version":2}