{"id":"37485d8e-6fd3-442b-86b3-0b9aa86d9b3d","arxiv_id":"2504.21581","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"LE-IRSTD, a YOLOv8-n variant using MBConv, partial convolution, variable-kernel convolution, and GSConv, reports a 2.6M-parameter model with mAP 95.4 on IRSTD-1K, outperforming listed SOTA methods.","lead":"This paper describes a lightweight infrared small-target detection model built on YOLOv8-n, combining several existing convolution blocks to cut parameters and FLOPs while reporting higher mAP on three public datasets. A smart generalist might read it to see how modular architecture tweaks and loss-weight tuning can move a benchmark number, and how fragile those claims are without error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA comparison depends on undisclosed hyperparameter selection: ablations in Tables III and VII end at exactly the final test mAP of 95.4, so without a validation-split protocol the headline numbers are not yet distinguishable from selection artifact.","rationale":"The reader's conditional verdict is justified. I looked for a more fundamental flaw — an architectural inconsistency or unsupported mathematical step — but the paper is a composition of known modules with plausible ablations, and the equations, while informal, are not the load-bearing issue. The central risk is experimental protocol: small absolute margins between LE-IRSTD and strong baselines, an undisclosed selection process, and no variance estimates. This is a fairness and internal-consistency concern, not a disagreement with field consensus. The provided code means a targeted check can settle it. I therefore keep the verdict at CONDITIONAL rather than moving to REJECT or ACCEPT.","tokens_in":17882,"tokens_out":3818,"duration_ms":42205,"concrete_test":"Ask the authors for the exact train/validation/test split manifests, or run the released code under a strict protocol: use only the 20% validation split for every ablation in Tables III and VII, freeze the selected configuration, and then evaluate once on the held-out test split. Repeat the final model with 3–5 seeds and report mean ± std. If the held-out test mAP falls materially below 95.4, or overlaps OSCAR's result within error bars, the headline SOTA claim is unsupported. Also verify that the baseline methods were retrained without any test-set-based selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that LE-IRSTD outperforms SOTA while being lighter — rests on the numbers in Table I. Those numbers are only meaningful if the configuration in the final row was chosen without using test labels. The paper never states this. Section IV.A.1 defines a 60/20/20 train/validation/test split, but Section IV.C.1 says IRSTD-1K was used for ablation studies, and Tables III and VII (width/depth/resolution and loss weights) both end at exactly the final Table I mAP of 95.4 on IRSTD-1K. If those ablations were evaluated on the test split, the 95.4 is a selection-maximized number, not an independent estimate. The margin over OSCAR is small (95.4 vs 95.0), so tuning alone could produce this gap. The same concern applies to the other datasets if test labels were inspected during development; at minimum the protocol is undisclosed. Additionally, all comparison methods were retrained for 300 epochs with their default parameters, while LE-IRSTD's own hyperparameters were apparently tuned, so the comparison is not apples-to-apples. The paper does provide code, so the issue is checkable rather than irreparable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LE-IRSTD, a lightweight infrared small target detection framework built on YOLOv8-n, where the backbone C2f modules are replaced by Mobile Inverted Bottleneck (MBConv) and Bottleneck Structure (BS) blocks, an Attention-based Variable Convolution Stem (AVCStem) with variable kernel convolution is introduced in the neck, and GSConv is used for downsampling. The loss is a weighted combination of BCE, CIoU, and a component the paper calls DFL. Experiments on IRSTD-1K, NUAA-SIRST, and NUDT-SIRST report 2.6M parameters, 7.4 GFLOPs, and mAP@50 of 95.4, 95.3, and 94.1, respectively, outperforming segmentation- and detection-based baselines including OSCAR, and the authors release the code. The paper argues that bounding-box regression is better suited to IRSTD than semantic segmentation because pixel-level IoU is overly sensitive to annotation ambiguity for tiny targets.","tokens_in":18154,"tokens_out":12587,"duration_ms":112664,"significance":"If the reported results hold, the paper would demonstrate that a compact YOLOv8 variant can beat much heavier segmentation-based IRSTD models with roughly 16x fewer parameters and 9x fewer FLOPs, which is a practically valuable result. The conceptual argument for bounding-box modeling of IRSTD is well motivated, and the lightweight design based on existing blocks is plausible. Concrete strengths include the public code release, the use of the mNoCoAP metric to compare across detection and segmentation paradigms, and systematic module-wise ablations. The central caveat is that the headline numbers currently rest on an undocumented hyperparameter-selection protocol and on single-run comparisons, so the quantitative claims are plausible but not yet established.","major_comments":[{"comment":"Section IV.A.1 defines a 60/20/20 train/validation/test split, but the paper never states that the architecture and loss-weight ablations in Tables III and VII were evaluated on the validation split rather than the test split. The final rows of both ablation tables report exactly the IRSTD-1K mAP of 95.4 that also appears in the final SOTA row of Table I, which is consistent with configuration selection being driven by test-set feedback. Because the reported margin over the runner-up (OSCAR, 95.0 in Table I) is only 0.4 mAP, tuning on test labels could plausibly account for the claimed advantage. The authors must state which split was used for all ablations and, if test labels were consulted during development, re-run the selection on validation data; otherwise the headline SOTA claim is not supported. In addition, all reported numbers appear to come from a single run, so the authors should provide multiple-seed means with standard deviations or confidence intervals for the central comparisons.","section":"IV.A.1, IV.C.1, Tables III and VII"},{"comment":"Section IV.B states that all comparison methods were retrained for 300 epochs with all other parameters at their default values, whereas LE-IRSTD's own configuration was selected through the ablations in Tables III-VII (loss weights 0.02/0.49/0.49, 6x expansion, 3x3 kernel, depths 1 and 2). Comparing a tuned model against default-configured baselines biases the comparison in favor of LE-IRSTD, so the claimed superiority over OSCAR and EFLNet is not an apples-to-apples result as presented. The authors should either give the baselines a comparable tuning budget, such as a small search over loss weights or resolution, or explicitly frame the comparison as tuned-model-versus-default-baselines and temper the SOTA claim accordingly.","section":"IV.B, IV.A.3"},{"comment":"Equation (17), labeled 'Dynamic Focal Loss (DFL),' reads L_DFL = -alpha(1-p)^gamma log(p), which is the standard Focal Loss formula; the Distribution Focal Loss used in YOLOv8 has a different form tied to soft labels over the box's corner distribution, and the accompanying text describing alpha as a class balance factor and gamma as a focusing factor confirms that the focal-loss formula, not DFL, is written down. If the implementation actually uses YOLOv8's DFL, the equation does not describe the implemented loss, and if focal loss is used, the method deviates from YOLOv8 and the loss-weight ablation should be interpreted accordingly. The notation also collides, since alpha denotes the VKConv scaling factor in Eq. (8), the CIoU weighting term in Eq. (15), and the focal balance factor in Eq. (17); please correct the formula, rename the symbols, and state which loss is actually implemented.","section":"III.C, Eq. (17)"},{"comment":"The ablation tables supporting the final configuration contain internal inconsistencies that need correction or explanation. In Table V, the SC row reports P=83.37, R=85.75, F1=82.77, but the harmonic mean of 83.37 and 85.75 is approximately 84.5, so the reported F1 is arithmetically impossible, and the VKConv row's F1 of 90.8 also differs from the harmonic mean of its P and R values (90.5). In Table II, row (c) reports F1=85.2 for P=88.2, R=81.7, whose harmonic mean is 84.8. In Table VII, the row (0.08, 0.46, 0.46) reports mAP 82.3, a large non-monotonic drop from the neighboring rows (90.7 and 87.5 mAP), which looks like a typo or an unstable run, and Section IV.C.6 states the optimum is lambda1=0.02, lambda2=0.02, lambda3=0.49, contradicting the table's lambda2=lambda3=0.49. These inconsistencies matter because Tables III and VII are the evidence that the reported configuration was chosen for accuracy rather than by chance.","section":"Tables II, V, VII; Section IV.C.6"}],"minor_comments":[{"comment":"The training description states 'an initial learning rate of 0.001 and a final learning rate of 0.5' with cosine annealing, which would increase the learning rate over training and is presumably a typo; it also lists 'momentum of 0.937', an SGD hyperparameter, while the stated optimizer is Adam. Please correct the hyperparameter description to match the implementation.","section":"IV.A.3"},{"comment":"The module is called 'Global Shuffle Convolution' in Sections III.A.3 and III.B.4 but 'Ghost Shuffle Convolution' in the Index Terms and in Contribution 3, and 'Attention-based Variable Convolution Stem' appears as 'Adaptive Vision Convolution Stem' in the Table II caption; Figure 3's caption also references an undefined 'Faststem' backbone and 'AVSFPN' neck. Please unify all module and component names.","section":"III.A.3, III.B.4, Table II caption, Fig. 3"},{"comment":"Equation (13) for the BCE loss has unbalanced parentheses as printed; the intended formula is L_BCE = -(y log(p(x)) + (1-y) log(1-p(x))).","section":"Eq. (13)"},{"comment":"The same method is cited as 'YOLO-SDLUWD' in the Introduction and as 'YOLO-SLWD' in Section II.B.2; please use a single consistent name.","section":"Introduction, II.B.2, Reference [17]"},{"comment":"The abstract and introduction claim real-time suitability, but no inference latency or FPS is reported anywhere in the experiments; either add latency measurements on the stated hardware or qualify the real-time claim.","section":"Abstract, I, IV.B"},{"comment":"Section IV.A.1 describes NUDT-SIRST as containing '1327 training and validation examples', which is ambiguous about the total image count used for the 60/20/20 split; please state the total size of each dataset and the exact split sizes.","section":"IV.A.1"},{"comment":"The text would benefit from a language pass: the abstract contains typos such as 'feature stackin' and 'lightwight', and sentences like 'while most existing methods model IRSTD as a semantic segmentation task, but they suffer from two critical drawbacks' are ungrammatical.","section":"Abstract, Index Terms"}],"recommendation":"major_revision","confidential_remarks":"This is an applied engineering contribution whose novelty lies in the combination of existing building blocks (MBConv, PConv, VKConv, GSConv) within a YOLOv8-n IRSTD detector; that is acceptable for an applications-oriented venue, but the introduction overstates the conceptual novelty. The main risk is the evaluation protocol: the final rows of the ablation tables reproduce the exact test mAP of 95.4 in the SOTA table, and the paper never discloses whether hyperparameter selection used the validation split, so the reviewer should weigh whether the authors' response resolves this before final acceptance. The mislabeled DFL equation (Eq. 17) also raises the question of whether the released code matches the paper; since code is available, this is verifiable, and I would encourage the editor to have that consistency checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this paper because it is a decent engineering contribution with public code, but treat the headline numbers with caution. The core claim—that a 2.6M-parameter, 7.4G-FLOP YOLOv8-n variant beats 42M-parameter segmentation models on three IRSTD benchmarks—is plausible, but the evaluation protocol as written does not yet support it.\n\nWhat is genuinely new is the specific combination of MBConv, PConv-based bottleneck, VKConv-based stem, and GSConv in a YOLOv8-n skeleton for infrared small target detection. The modules are all known, and the paper borrows heavily (MBConv from MobileNets, PConv from FasterNet, GSConv from YOLO neck literature, VKConv is essentially LDConv), but the assembled architecture and the AVCStem are new to this task. The authors provide code and detailed training settings, so the empirical claims are checkable. That is real value.\n\nThe soft spot is the one the stress-test flags, and it holds up on reading. Section IV.A.1 defines a 60/20/20 train/val/test split, but Section IV.C.1 says IRSTD-1K was used for ablation studies. Tables III and VII (width/depth/resolution and loss weights) both end at exactly the final test mAP of 95.4 on IRSTD-1K. The paper never states these ablations used the validation split. If they used test labels for hyperparameter selection, the 95.4 is a selection artifact, and the margin over OSCAR (95.4 vs 95.0) is too small to rule that out. The same concern applies to the other two datasets if test labels were inspected during development. This is the main issue, and it is fixable: clarify the protocol, or re-run ablations on the validation split and report test results only for the final configuration.\n\nSecondary issues: no error bars or multiple seeds; all comparison methods were retrained with default hyperparameters while LE-IRSTD's were tuned, so the comparison is not apples-to-apples; Eq. 17 is focal loss, not DFL; and the text in Section IV.C.6 gives lambda1=0.02, lambda2=0.02, lambda3=0.49 while Table VII's best row is 0.02, 0.49, 0.49. Minor typos aside, the missing validation split is the load-bearing one.\n\nWho this is for: anyone working on lightweight IRSTD or edge detection will find the architecture and ablations useful, and the code is worth looking at. It deserves a serious referee—the engineering is real and the problems are correctable. My recommendation: send it to review, but the revision must clarify the validation protocol and ideally add multi-seed results. If the protocol turns out to be test-based, the SOTA claim should be downgraded to 'competitive' rather than 'superior.'","headline":"A competent lightweight IRSTD detector with public code, but the SOTA claim rests on an undisclosed validation protocol that the paper needs to clarify.","tokens_in":18700,"tokens_out":3194,"would_cite":false,"duration_ms":29928,"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":"The paper argues that a lightweight YOLOv8-n-based detector with four efficiency modules matches or beats segmentation-based methods on infrared small target detection at a fraction of the compute.","keywords":["infrared small target detection","lightweight network","YOLOv8","Mobile Inverted Bottleneck Convolution","Partial Convolution","Variable Kernel Convolution","GSConv","real-time detection"],"falsifier":"Run a validation-only hyperparameter search over the loss weights and MBConv width settings, lock the chosen values, and then evaluate on the test split exactly once; if the resulting IRSTD-1K mAP falls below the reported 95.4 or below OSCAR's 95.0, the claimed state-of-the-art result would not survive an unbiased protocol.","tokens_in":17683,"feed_emoji":"🎯","tokens_out":8049,"duration_ms":75041,"temperature":0.7,"pith_summary":"The paper tries to show that infrared small target detection (IRSTD) can be reformulated as bounding-box detection rather than pixel-level segmentation without sacrificing accuracy. It introduces LE-IRSTD, built on the YOLOv8-n detector, whose backbone and neck replace heavy standard convolutions with Mobile Inverted Bottleneck blocks, partial-convolution bottleneck blocks, an attention-based variable-kernel convolution stem, and channel-shuffle convolution. On three public IRSTD benchmarks the model reports mAP@50 of 95.4, 95.3, and 94.1 with 2.6M parameters and 7.4 GFLOPs, surpassing several larger segmentation and detection models. The significance is practical: if the claim holds, real-time infrared surveillance and targeting systems can get high detection accuracy on small, dim targets without the compute budget segmentation methods demand.","feed_headline":"2.6M-parameter YOLO variant tops IRSTD benchmarks at 95.4 mAP","feed_subtitle":"Small-target detection from bounding boxes, not segmentation, keeps accuracy high at 7.4 GFLOPs.","key_machinery":"The load-bearing machinery is the LE-IRSTD architecture itself, a YOLOv8-n derivative with four substitutions. MBConvblock expands channels $6\\times$ with a $1\\times 1$ convolution, applies a $3\\times 3$ depthwise convolution, adds CBAM attention and dropout, and uses a residual connection when channels match; it replaces repeated C2f bottlenecks to cut depth. BSblock uses Partial Convolution (PConv) on one quarter of the channels, concatenates the untouched channels, passes through an MLP with dropout, and residual-adds; the paper says this cuts memory access to 25% and FLOPs to $1/16$ of standard convolution at ratio $r = 1/4$. AVCStem fuses parallel spatial and channel branches and uses VKConv to learn per-pixel offsets (scaled by $\\alpha = 0.1$) that deform the $5\\times 5$ sampling grid to match target shape. GSConv runs a $3\\times 3$ convolution, then a $3\\times 3$ depthwise convolution, concatenates, and channel-shuffles the result. The loss is a weighted sum $L_{\\mathrm{all}} = 0.02 L_{\\mathrm{BCE}} + 0.49 L_{\\mathrm{CIoU}} + 0.49 L_{\\mathrm{DFL}}$, deliberately down-weighting classification because the task has one class.","core_discovery":"On its own terms, the paper's central claim is that the accuracy/efficiency trade-off in IRSTD is not fixed: a careful choice of lightweight blocks can improve both at once. LE-IRSTD combines four modules — MBConvblock for expand-then-depthwise feature extraction, BSblock which applies partial convolution to a quarter of the channels and preserves the rest, AVCStem which adapts convolution kernel shapes via variable kernel convolution (VKConv) to fit irregular small targets, and GSConv which interleaves standard and depthwise features with channel shuffle. With these, the model reports 89.8 precision, 91.2 recall, 90.5 F1, and 95.4 mAP@50 (mean average precision at IoU threshold 0.5) on IRSTD-1K, 95.3 mAP on NUAA-SIRST, and 94.1 mAP on NUDT-SIRST, while reducing parameters from 3.0M to 2.6M and FLOPs from 8.2G to 7.4G relative to the YOLOv8-n baseline. The paper further claims that the bounding-box formulation avoids the IoU 'cliff drop' that pixel-level annotation ambiguity causes for small targets.","pith_inferences":["A natural extension is to test the same four-module recipe on non-infrared small-object benchmarks such as aerial, drone, or maritime imagery; the paper's mechanism is generic, so if the gains persist the method is a general small-target efficiency recipe rather than an infrared-specific one.","The paper's use of mNoCoAP suggests a testable prediction: bounding-box detectors trained with heavy regression loss will show smaller mNoCoAP-to-mAP gaps than segmentation-based methods, because they optimize centroid localization directly.","The hyperparameter search over loss weights (Table VII) could be rerun under strict validation-only selection; confirming the reported 95.4 mAP under that protocol would settle whether the state-of-the-art claim depends on test-set feedback.","The PConv ratio $r = 1/4$ and MBConv expansion factor $6$ are reported as chosen values, not proven optima; a validation-set sensitivity sweep over these two numbers would reveal how sharply performance depends on them."],"forward_implications":["If LE-IRSTD's numbers are reproducible, bounding-box detection becomes a viable default paradigm for IRSTD, shifting the field's focus from segmentation refinement to detector efficiency.","The 7.4 GFLOPs and 2.6M-parameter budget makes real-time onboard deployment plausible on edge GPUs or embedded devices, which the much heavier segmentation baselines cannot offer.","The loss-weighting result implies that for single-class small-target detection, allocating nearly all loss weight to box regression and localization (CIoU and DFL) rather than classification improves mAP.","The reported mNoCoAP improvements (e.g., 87.2 versus 85.0 for OSCAR on IRSTD-1K) indicate that the gain is not only box overlap but also target contrast and centroid localization, a metric designed to be paradigm-independent.","The four modules are generic enough that swapping C2f bottlenecks for MBConv/BS blocks and standard neck convolutions for GSConv could reduce FLOPs in other YOLO-family small-object detectors."],"supporting_citations":[{"why":"Supplies the IRSTD-1K dataset, the primary benchmark for the main comparison and all ablations.","marker":"[8]"},{"why":"Supplies the NUAA-SIRST dataset, one of the three evaluation benchmarks.","marker":"[36]"},{"why":"Supplies the NUDT-SIRST dataset and the dense-nested-attention baseline context.","marker":"[35]"},{"why":"Provides OSCAR, the strongest detection-based baseline (95.0 mAP on IRSTD-1K), and the mNoCoAP metric that LE-IRSTD improves on.","marker":"[16]"},{"why":"Provides EFLNet, a detection-based baseline whose high complexity the paper contrasts with LE-IRSTD's efficiency.","marker":"[19]"},{"why":"Provides SCTransNet, a strong segmentation-based baseline that LE-IRSTD outperforms in mAP and efficiency.","marker":"[43]"},{"why":"Supplies the variable-kernel/deformable convolution mechanism (VKConv) used inside AVCStem.","marker":"[34]"},{"why":"Supplies the CBAM attention module used inside MBConvblock.","marker":"[32]"},{"why":"Provides RetinaNet, a detection baseline with FPN and focal loss used in the comparison.","marker":"[31]"},{"why":"Provides AMFLW-YOLO, a lightweight detection baseline at 2.0M parameters and 4.6 GFLOPs, making it a direct efficiency competitor.","marker":"[46]"}],"fun_headline_variants":["2.6M-param LE-IRSTD hits 95.4 mAP on IRSTD-1K","Lightweight YOLO variant cuts FLOPs to 7.4G, raises mAP to 95.4","Bounding-box IRSTD: 95.4 mAP with 2.6M parameters","MBConv, AVCStem, GSConv make IRSTD lean and accurate","Small-target detection hits 95.4 mAP at 7.4 GFLOPs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the model's settings, especially the loss weights $\\lambda_1 = 0.02$, $\\lambda_2 = 0.49$, $\\lambda_3 = 0.49$, were chosen on a validation split and not by looking at test-set labels; the paper does not state this, and the ablation tables report the same final mAP as the main comparison.","fun_headline_variants_meta":{"raw":{"variants":["2.6M-param LE-IRSTD hits 95.4 mAP on IRSTD-1K","Lightweight YOLO variant cuts FLOPs to 7.4G, raises mAP to 95.4","Bounding-box IRSTD: 95.4 mAP with 2.6M parameters","MBConv, AVCStem, GSConv make IRSTD lean and accurate","Small-target detection hits 95.4 mAP at 7.4 GFLOPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000883,"raw_usage":{"total_tokens":3907,"prompt_tokens":1132,"completion_tokens":2775,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":748,"completion_tokens_details":{"reasoning_tokens":2644}},"tokens_in":748,"tokens_out":2775,"duration_ms":21542,"temperature":1.0,"reasoning_tokens":2644,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:58:37.426916+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a validation-only hyperparameter search over the loss weights and MBConv width settings, lock the chosen values, and then evaluate on the test split exactly once; if the resulting IRSTD-1K mAP falls below the reported 95.4 or below OSCAR's 95.0, the claimed state-of-the-art result would not survive an unbiased protocol.","supporting_citations":[{"cited_title":"ISNET: Shape matters for infrared small target detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the IRSTD-1K dataset, the primary benchmark for the main comparison and all ablations."},{"cited_title":"Asymmetric contextual modulation for infrared small target detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the NUAA-SIRST dataset, one of the three evaluation benchmarks."},{"cited_title":"Dense nested attention network for infrared small target detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the NUDT-SIRST dataset and the dense-nested-attention baseline context."},{"cited_title":"One-stage cascade refinement networks for infrared small target detection,","cited_arxiv_id":null,"evidence_quote":"Provides OSCAR, the strongest detection-based baseline (95.0 mAP on IRSTD-1K), and the mNoCoAP metric that LE-IRSTD improves on."},{"cited_title":"Eflnet: Enhancing feature learning network for infrared small target detection,","cited_arxiv_id":null,"evidence_quote":"Provides EFLNet, a detection-based baseline whose high complexity the paper contrasts with LE-IRSTD's efficiency."},{"cited_title":"Sctransnet: Spatial-channel cross transformer network for infrared small target detection,","cited_arxiv_id":null,"evidence_quote":"Provides SCTransNet, a strong segmentation-based baseline that LE-IRSTD outperforms in mAP and efficiency."},{"cited_title":"Ldconv: Linear deformable convolution for improving convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the variable-kernel/deformable convolution mechanism (VKConv) used inside AVCStem."},{"cited_title":"Cbam: Convolutional block attention module,","cited_arxiv_id":null,"evidence_quote":"Supplies the CBAM attention module used inside MBConvblock."},{"cited_title":"Focal loss for dense object detection,","cited_arxiv_id":null,"evidence_quote":"Provides RetinaNet, a detection baseline with FPN and focal loss used in the comparison."},{"cited_title":"Amflw-yolo: A lightweight network for remote sensing image detection based on attention mechanism and multi-scale feature fusion,","cited_arxiv_id":null,"evidence_quote":"Provides AMFLW-YOLO, a lightweight detection baseline at 2.0M parameters and 4.6 GFLOPs, making it a direct efficiency competitor."}],"review_version":1}