{"id":"a3a522a9-ed45-4c9a-a292-77ff6d160861","arxiv_id":"2412.04931","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DEYOLO fuses RGB and infrared features with dual channel and spatial enhancement modules and reports improved object detection accuracy on the M3FD and LLVIP datasets.","lead":"The paper presents DEYOLO, a detection network that combines ordinary visible-light photos with infrared images to find objects in dark or smoky scenes. It adds two attention-style modules that let each image type sharpen the other's features, and reports higher detection scores than several existing methods on two public datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 4 refutes the abstract's 'clear margin' claim: single-modality YOLOv7 (visible) scores 90.4/61.3, above DEYOLO-n's 86.6/58.9, and no baseline is run on the held-out test split.","rationale":"The paper's strongest asset is its detailed ablation (Table 1), which is monotone and supports each module; the code link is a real plus. The central problem is not the architecture but the comparison. The reader's weakest assumption (validation-only baselines) is correct and is compounded by an internal numerical contradiction: Table 4 shows a standard single-modality YOLOv7 at 90.4/61.3, above DEYOLO-n at 86.6/58.9. Since the abstract promises a 'clear margin' over SOTA detection algorithms, the paper as written cannot support that promise for the n variant. The l variant does top the table, but only by 0.8 mAP50 over YOLOv7, and with no parameter-matched baselines or test-split numbers for any competitor, so the margin is unverified. I therefore keep the reader's conditional verdict: the claim can probably be rescued by re-running all comparisons on the held-out test split and reporting parameter counts, but it is not currently established.","tokens_in":10965,"tokens_out":7917,"duration_ms":78868,"concrete_test":"Rerun all single-modality baselines in Table 4 (at least YOLOv7 and YOLOv8-l) and all DEYOLO variants on the same 3:1:1 test split used for the one reported test number, with identical 800-epoch training and the same hyperparameter selection protocol, reporting test mAP50/mAP50-95 and parameter counts. If YOLOv7's test mAP50 is above DEYOLO-n's 85.7 (the reported test value), the Abstract's claim is directly contradicted for the n variant; if a parameter-matched YOLOv8-l or YOLOv7-l baseline closes the gap to DEYOLO-l, the 'clear margin' claim must be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is an internal inconsistency in the paper's headline comparison. In Table 4, DEYOLO-n (ours) scores 86.6 mAP50 and 58.9 mAP50-95, while single-modality YOLOv7 trained on visible images scores 90.4 and 61.3, and YOLOv8-l scores 88.3 and 61.8. The abstract claims 'our approach outperforms SOTA object detection algorithms by a clear margin' and Section 4.4 asserts 'DEYOLO outperforms all single-modality object detection models,' but Table 4 shows a standard single-modality detector beating DEYOLO-n on both metrics. Only DEYOLO-l (91.2/66.3) tops the table, and that comparison lacks parameter-matched baselines, per-model parameter counts, and any baseline evaluated on the held-out 3:1:1 test split (the paper reports test mAP only for DEYOLO). Thus the central 'clear margin' claim is not supported by the paper's own results as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DEYOLO, a YOLOv8-based RGB-infrared object detector. It introduces a dual semantic enhancing channel weight assignment module (DECA), a dual spatial enhancing pixel weight assignment module (DEPA), and a bi-directional decoupled focus module in the backbone. The method is evaluated on M3FD and LLVIP under a self-defined 3:1:1 split, with additional experiments on KAIST, and the paper claims that DEYOLO outperforms state-of-the-art object detectors and fusion-and-detection methods by a clear margin.","tokens_in":11200,"tokens_out":6716,"duration_ms":63747,"significance":"If the empirical claims were fully supported, the paper would make a useful contribution: detection-centered cross-modality fusion is a practical direction, the proposed modules are simple and plausibly plug-and-play, and the ablation in Table 1 shows consistent gains from each component. The code release is also a concrete reproducibility asset. However, the current evidence does not support the stated 'clear margin' over single-modality state-of-the-art detectors, and the evaluation protocol needs substantial strengthening before the headline claim can be accepted.","major_comments":[{"comment":"The headline claim that DEYOLO 'outperforms SOTA object detection algorithms by a clear margin' is contradicted by the paper's own Table 4. DEYOLO-n obtains 86.6 mAP50 / 58.9 mAP50-95, while single-modality YOLOv7 (visible) obtains 90.4 / 61.3 and YOLOv8l (visible) obtains 88.3 / 61.8. Only DEYOLO-l (91.2 / 66.3) tops the table, and for that comparison the table does not report parameter-matched baselines, per-model parameter counts, or a YOLOv8l cross-modality row. Please either add the missing baselines and complexity measures or revise the 'clear margin' and 'outperforms all single-modality' claims.","section":"Abstract; Section 4.4, Table 4"},{"comment":"All headline numbers in Tables 4-5 are mAP on a self-defined 3:1:1 validation split, and the same split was used to select DEPA's kernel size (Table 2), DECA's layer type and count (Table 3), and module placement. The test split is mentioned only in one sentence reporting DEYOLO's test mAP, with no baseline evaluated on it. This makes the claimed margins selection-dependent; please report test-split results for all compared methods or use the official train/test splits.","section":"Section 4.2-4.4, Tables 2-5"},{"comment":"The ablation baseline is 'the baseline network trained merely by visible images' (80.8 / 54.3), while the DEYOLO rows use both modalities. Since the improvements co-occur with the addition of the infrared stream, Table 1 does not isolate the contribution of DECA, DEPA, and the bi-direction focus from the benefit of simply having a second modality. Please add a cross-modality baseline without the proposed modules, such as YOLOv8n with a four-channel input or with two-branch late fusion, and report its mAP in Table 1.","section":"Section 4.3, Table 1"},{"comment":"The margin over fusion-and-detection methods is not 'clear' for the small model: on LLVIP, DEYOLO-n leads the next best method (SeAFusion) by only 0.6 mAP50, and on M3FD the stated minimum gains of 5.4 / 3.1 points are computed only against the fusion-and-detection baselines, not against the single-modality detectors in Table 4. In addition, the fusion baselines are not retrained on the same 3:1:1 split or with the same protocol, so the comparison may conflate detector training choices with fusion quality. Please clarify the training protocol for all baselines and temper the margin claims accordingly.","section":"Section 4.4, Table 5"}],"minor_comments":[{"comment":"The caption contains the typo 'U2Fsuion' for 'U2Fusion'; please fix it.","section":"Section 1, Fig. 1 caption"},{"comment":"The convolution operations conv1 and conv2 in Eq. (7) are not defined in the method text; the default kernel size is only inferable from Table 2. Please specify these operations explicitly.","section":"Section 3.2, Eq. (7)"},{"comment":"The phrase 'an element-wise multiplication is applied on the result of each other' is unclear; please state explicitly that the two convolved features are multiplied element-wise.","section":"Section 3.2, Eq. (6)"},{"comment":"The text says the authors use Log Average Miss Rate (LAMR) as an evaluation metric, but no LAMR values are reported anywhere in the paper; please add them or remove the sentence.","section":"Section 4.2"},{"comment":"The YOLOv8l row is missing a cross-modality entry, and no parameter counts or FLOPs are provided for any model; please complete the table to make the comparisons interpretable.","section":"Table 4"},{"comment":"Figure 3 labels parts of DECA as 'Linear' while the text describes convolutions (Eqs. (2)-(4)); please align the figure with the text.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"I concur with the conditional assessment: the architecture and ablations are valuable, but the headline comparison is overstated. The revision should focus on baseline fairness, parameter-matched comparisons, and test-split reporting rather than new algorithmic development."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know before reading: the architecture is plausible and the ablation is clean, but the central claim is overstated. The paper's own Table 4 shows single-modality YOLOv7 beating DEYOLO-n on both M3FD metrics, which contradicts the abstract's 'clear margin' promise.\n\nWhat's new: DECA and DEPA are a channel-attention and spatial-attention pair with a crossed dual-enhancement scheme—each modality's weights multiply onto the other modality's features, and the fused result is enhanced as well. The bi-direction decoupled focus is a modest twist on YOLOv5's Focus, but it's clearly described. The ablation in Table 1 shows each module adds a few points, which supports the architecture's basic value. The writing is readable, and the code link is a plus.\n\nThe soft spots, in proportion: the headline claim is the big one. Section 4.4 says DEYOLO outperforms all single-modality detectors, but Table 4 has DEYOLO-n at 86.6/58.9 versus YOLOv7-visible at 90.4/61.3 and YOLOv8-l-visible at 88.3/61.8. Only DEYOLO-l tops the table, and that comparison has no parameter-matched baselines. The fusion-and-detection comparison in Table 5 is kinder, but those baselines are fusion models retrained with YOLOv8, which is a low bar.\n\nSecond, the evaluation protocol is fragile. The authors re-split into 3:1:1, used the validation split to pick kernel sizes and layer counts (Tables 2 and 3), then report validation mAP as the headline. The test numbers appear in one sentence, with no baseline evaluated on that test split. If the validation-tuned margins shrink on a truly held-out set, the 'clear margin' claim fails even if the modules still help.\n\nAlso minor: the related-work claim that existing works 'only focus on image fusion' ignores the direct RGB-IR detection networks the paper itself compares against on KAIST. That overstatement should be softened.\n\nThe paper is still worth engaging with. The modules are well-motivated, the ablation is consistent, and the authors show the KAIST numbers even though they don't beat SOTA there, which is honest. With a corrected abstract, baselines on a proper test split, and parameter-matched comparisons at each scale, this could be a modest, useful contribution.\n\nRecommendation: send it to peer review with major revision. A serious referee should require the claims to match the tables and the evaluation to be redone on a held-out test split. If the margins survive that, accept as an incremental improvement.","headline":"A sensible incremental RGB-IR detection architecture with a clean ablation, but the headline 'clear margin' claim is refuted by the paper's own Table 4.","tokens_in":11794,"tokens_out":3319,"would_cite":false,"duration_ms":29517,"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 DEYOLO, a YOLOv8-based detector, outperforms state-of-the-art single-modality and fusion-and-detection methods on RGB-infrared object detection in poor illumination by fusing modalities in feature space with…","keywords":["cross-modality object detection","RGB-infrared fusion","feature-level fusion","dual enhancement","channel attention","spatial attention","low-light detection","YOLOv8"],"falsifier":"A reader could settle it by taking the same 3:1:1 split, retraining each fusion-and-detection baseline and DEYOLO under identical schedules, and reading mAP50 on the test portion: if any baseline matches or beats 85.7 on M3FD or 96.4 on LLVIP, the paper's central margin claim fails.","tokens_in":10756,"feed_emoji":"🌙","tokens_out":7599,"duration_ms":71397,"temperature":0.7,"pith_summary":"The paper tries to establish that fusing visible and infrared images inside a YOLO detector, rather than fusing the images first and then detecting, improves object detection in low-light scenes, and that its specific dual-enhancement modules reduce the mutual interference between the two modalities. If that is right, poor-illumination detection can be improved by a plug-in feature-space fusion that does not require a separate image-fusion stage. The claim is supported by experiments on the M3FD and LLVIP datasets, where the proposed DEYOLO reports higher mAP than single-modality detectors and fusion-and-detection pipelines.","feed_headline":"RGB-IR fusion inside a YOLO detector beats fusion-first pipelines","feed_subtitle":"Channel and spatial dual-enhancement modules fuse visible and infrared features to lift mAP on low-light datasets.","key_machinery":"The machinery is a pair of dual-enhancement modules inserted between the YOLOv8 backbone and the detection heads. DECA computes channel weights: one from each modality and one from the concatenated mixed feature, multiplies them to enhance the mixed feature, then multiplies each original modality feature by the other modality's enhanced weight. DEPA does the same in the spatial dimension with convolutional pixel weights, and the enhanced RGB and IR features are added for detection. A third component, the bi-directional decoupled focus, splits the early feature map into two pixel groups for horizontal and vertical downsampling and concatenates them, enlarging the receptive field while preserving surrounding information. The dual-enhancement loop, where single modality enhances fusion and fusion enhances single modality, is what carries the claimed reduction of cross-modality interference.","core_discovery":"The central discovery claimed is that cross-modality RGB-infrared object detection benefits from a dual-enhancement fusion in feature space: a channel-weight module (DECA) and a spatial-weight module (DEPA) each first use single-modality information to sharpen a mixed two-modality feature, then use the enhanced mixed feature to sharpen each modality again, and this two-way exchange reduces interference between RGB and IR. Adding a bi-directional decoupled focus to the backbone enlarges the receptive field in horizontal and vertical directions. On M3FD and LLVIP, the authors report that this combination outperforms state-of-the-art single-modality detectors and fusion-and-detection methods by a clear margin, with DEYOLO-l reaching 91.2 mAP50 on M3FD and DEYOLO-n 96.8 mAP50 on LLVIP.","pith_inferences":["If the validation gains survive on a true held-out test set, the practical takeaway is that feature-level, detection-oriented fusion is a cheaper and more direct route than image-fusion-then-detect pipelines for low-light RGB-IR systems.","The same dual-enhancement pattern could be tested on other paired inputs, such as RGB-depth or multi-spectral images, where one modality is reliable and the other is complementary.","A reader should treat the reported margins as validation-split estimates: the paper selects DEPA kernel size and DECA depth on the same split, so independent replication on a different split or dataset is the natural next check."],"forward_implications":["On M3FD, adding DECA, DEPA, and the bi-directional focus lifts YOLOv8n from 80.8 to 86.6 mAP50; on LLVIP, DEYOLO-n reaches 96.8 mAP50, above all listed fusion-and-detection baselines.","A detector can exploit infrared information without a separate image-fusion network, so training and inference become a single end-to-end detection model.","DECA and DEPA are described as plug-and-play, meaning other detection backbones could adopt the same dual-enhancement fusion pattern.","The reported KAIST results suggest the approach also applies to RGB-thermal detection, though not to state-of-the-art level there.","The dual-enhancement logic implies that mutual interference between modalities can be handled in feature space, not only at the image level."],"supporting_citations":[{"why":"Supplies the YOLOv8 backbone and baseline that DEYOLO builds on; all ablation and comparison numbers are anchored to this detector.","marker":"[12]"},{"why":"Provides the M3FD dataset used for ablations and the main M3FD comparison.","marker":"[16]"},{"why":"Provides the LLVIP low-light visible-infrared dataset used for the second main comparison.","marker":"[10]"},{"why":"Supplies the squeeze-and-excitation channel-weight idea that DECA's channel weight extraction adapts.","marker":"[8]"},{"why":"Provides the focus and slicing downsampling operation that the bi-directional decoupled focus modifies.","marker":"[11]"},{"why":"Provides depthwise convolution used inside the bi-directional decoupled focus and in DECA's weight extraction.","marker":"[4]"},{"why":"One of the fusion-and-detection baselines DEYOLO is compared against for the core superiority claim.","marker":"[24]"},{"why":"Another fusion-and-detection baseline whose unsupervised fusion pipeline is a representative alternative DEYOLO must beat.","marker":"[31]"}],"fun_headline_variants":["Dual-enhancement RGB-IR fusion lifts YOLO low-light detection","DEYOLO's two-way feature enhancement fuses RGB and IR for detection","Channel and spatial dual-enhancement modules improve cross-modality detection","Bi-directional focus and dual-enhancement fusion boost RGB-IR detection","Feature-space dual enhancement reduces RGB-IR interference in YOLO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that the validation split used for all headline comparisons is an honest measure of detection quality, even though that same split was used to choose module hyperparameters and no baseline was scored on the held-out test portion.","fun_headline_variants_meta":{"raw":{"variants":["Dual-enhancement RGB-IR fusion lifts YOLO low-light detection","DEYOLO's two-way feature enhancement fuses RGB and IR for detection","Channel and spatial dual-enhancement modules improve cross-modality detection","Bi-directional focus and dual-enhancement fusion boost RGB-IR detection","Feature-space dual enhancement reduces RGB-IR interference in YOLO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000866,"raw_usage":{"total_tokens":3801,"prompt_tokens":1043,"completion_tokens":2758,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":2659}},"tokens_in":659,"tokens_out":2758,"duration_ms":19267,"temperature":1.0,"reasoning_tokens":2659,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:08:07.185832+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle it by taking the same 3:1:1 split, retraining each fusion-and-detection baseline and DEYOLO under identical schedules, and reading mAP50 on the test portion: if any baseline matches or beats 85.7 on M3FD or 96.4 on LLVIP, the paper's central margin claim fails.","supporting_citations":[{"cited_title":"https://github.com/ultralytics/ultralytics (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the YOLOv8 backbone and baseline that DEYOLO builds on; all ablation and comparison numbers are anchored to this detector."},{"cited_title":"In: Proceedings of the IEEE/CVF international con- ference on computer vision","cited_arxiv_id":null,"evidence_quote":"Provides the LLVIP low-light visible-infrared dataset used for the second main comparison."},{"cited_title":"https://doi.org/10.5281/ zenodo.3908559, https://github.com/ultralytics/yolov5","cited_arxiv_id":null,"evidence_quote":"Provides the focus and slicing downsampling operation that the bi-directional decoupled focus modifies."},{"cited_title":"In: Proceedings of the 30th ACM International Conference on Multimedia","cited_arxiv_id":null,"evidence_quote":"One of the fusion-and-detection baselines DEYOLO is compared against for the core superiority claim."},{"cited_title":"IEEE Transactions on Pattern Analysis and Machine Intelligence 44(1), 502–518 (2020)","cited_arxiv_id":null,"evidence_quote":"Another fusion-and-detection baseline whose unsupervised fusion pipeline is a representative alternative DEYOLO must beat."}],"review_version":1}