{"id":"4dfc266b-33e8-4e87-b773-cfc0b39ac0b9","arxiv_id":"2505.12820","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors propose an unfused pyramid, a scaled soft nearest-neighbor upsampling, an adaptive downsampling block, and a lightweight convolution variant, and report improved real-time detection accuracy on VOC and COCO.","lead":"This paper argues that the standard fused feature pyramid neck in object detectors hurts accuracy by mixing misaligned features, and proposes four modifications to YOLO-style necks. The combined system reports small accuracy gains on Pascal VOC and MS COCO, but without code or error bars the size of the real effect is hard to verify.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SNI is a global scalar after nearest-neighbor upsampling, not a spatial alignment operation; any gain is absorbable by the following conv-BN, so the experiments do not establish the misalignment mechanism.","rationale":"The reader's weakest assumption concerned unmatched parameter and FLOP counts in Tables 4-6. That is a valid threat to the SOTA claim. However, the most load-bearing issue lies one level deeper: the signature SNI component, as specified in Eq. (1), is not a soft interpolation at all. It is a global gain applied after nearest-neighbor upsampling. Because the fusion modules begin with conv-BN stacks, the gain can be absorbed into the following weights, so the function class is unchanged. The ablation evidence therefore cannot distinguish feature alignment from a scalar-induced change in optimization dynamics. This does not automatically falsify the empirical gains; it means the paper's central mechanistic interpretation and the claim that SNI 'confirms' misalignment are not supported by the reported experiments. The right remedy is an additional control (a reparameterization test plus an alpha sweep), not an outright rejection. I thus keep the reader's CONDITIONAL verdict; the added concern is a more precise missing control than the one the reader highlighted.","tokens_in":14230,"tokens_out":8509,"duration_ms":92160,"concrete_test":"Run YOLOv5-n-PANet on VOC 07+12 with: (a) standard nearest-neighbor upsampling, (b) SNI α=0.5 from Eq. (1), and (c) standard upsampling but with the first convolution after each concat initialized with weights divided by 0.5 on the upsampled-branch input channels so the initial function matches SNI. If (b) and (c) converge to the same AP, SNI is a reparameterization or optimization artifact; if (c) matches (a), the reported gain is entirely a training-dynamics effect. Also sweep α ∈ {0.25, 0.75, 1.0, 2.0}; if Eq. (1)'s value is not uniquely best, the resolution-ratio formula is not load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (1) defines SNI as Y = α·f(X) with α = ResolutionX/ResolutionY, i.e., a single fixed scalar multiplied onto the whole nearest-neighbor-upsampled feature map. This is spatially uniform: it multiplies every pixel and every channel by the same factor, so it cannot change which spatial locations are aligned or misaligned. It is also reparameterizable in the necks used. In the PANet-style necks of the six YOLO baselines, the upsampled branch is concatenated (or added) with the lateral branch and then passed through a convolution followed by BatchNorm. A constant scaling of one input branch is exactly equivalent to scaling the corresponding input-channel weights of that first convolution by 1/α, with BatchNorm absorbing the remaining scale. The set of representable functions is therefore unchanged; only initialization and gradient scales differ. Consequently the consistent AP gains in Table 1, Part II do not confirm 'the presence of features misalignment' asserted in Sec. 3.2; they are equally consistent with a generic optimization or regularization effect from changing the branch's initial scale. The paper provides no sensitivity study for α, no learned-scalar baseline, and no analysis ruling out reparameterization. The central mechanistic claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that point-to-point fusion of feature maps from different pyramid levels in FPN-style necks causes feature misalignment, and that this hurts real-time object detectors. To address it, the authors propose an independent hierarchy pyramid (IHP) neck that removes fusion, a soft nearest neighbor interpolation (SNI) that down-weights upsampled high-level features by a fixed scalar alpha = ResolutionX / ResolutionY, a features adaptive selection downsampling method (ESD), and a lightweight convolution module (GSConvE). These components are combined into a 'secondary features alignment' (SA) solution and evaluated on Pascal VOC and MS COCO across six YOLO baselines, with claims of state-of-the-art results. The paper also reports a comparison of confidence thresholds to motivate a more 'realistic' evaluation protocol.","tokens_in":14488,"tokens_out":5169,"duration_ms":49723,"significance":"The empirical breadth is a genuine strength: the IHP, SNI, ESD, and GSConvE are tested across six YOLO baselines, the SNI gains are consistent, and the architecture table plus the promise of released code would aid reproducibility. If the claimed effects are real, a zero-cost scalar multiplier on upsampled features plus simple downsampling and convolution changes would be a practically useful recipe for real-time detectors. However, the central mechanistic claim about feature misalignment is not established by the experiments as designed, and the headline comparisons are not parameter- or FLOP-matched. The paper would be significantly strengthened by control experiments that separate capacity effects from the proposed mechanisms.","major_comments":[{"comment":"The 'soft nearest neighbor interpolation' is a single scalar alpha = ResolutionX / ResolutionY multiplied onto the entire upsampled feature map. Because every spatial location and every channel receives the same factor, it cannot change which features are spatially aligned. In the PANet-style necks used here, the upsampled branch is concatenated or added to a lateral branch and then passed through a convolution followed by BatchNorm; a constant scaling of one input branch is exactly equivalent to rescaling the corresponding weights of that convolution, with BatchNorm absorbing the scale. The representable function class is therefore unchanged, so the consistent gains in Table 1 Part II do not by themselves confirm the existence of feature misalignment. The authors should add a learned-scalar baseline (e.g., a per-channel or global learned alpha), a sensitivity study over alpha, and a reparameterization check to show that the gain is not simply an initialization or optimization effect.","section":"Sec. 3.2, Eq. (1)"},{"comment":"The headline state-of-the-art comparisons are not matched in parameters or FLOPs. For example, SYOLO has 57.2M parameters and 142.5 GFLOPs, while YOLOv8 has 43.7M and 165.2 GFLOPs; SYolo-s has 6.6M parameters while YOLOv8-n has 3.2M. Differences of this size mean the accuracy gaps could be due to extra capacity rather than the proposed SNI, ESD, and GSConvE components. The authors should report matched-capacity baselines, or ablations in which the SA components are swapped into the exact same backbone-neck-head with the same parameter and FLOP budget.","section":"Sec. 4.3, Tables 4-6"},{"comment":"The reported gain for Yolov3-t-esd-I is 11.2 AP and 11.2 AP50 (from 58.0 to 69.2 AP50 and 27.3 to 38.5 AP), while the parameter count grows from 8.71M to 11.8M and FLOPs from 13.0G to 22.4G. This single outlying result is far larger than all other ESD gains and is not discussed in the text. It suggests either a configuration difference, an extra-capacity effect, or a possible reporting error. The aggregate claim that ESD improves accuracy is not supported until this row is explained or corrected.","section":"Table 1, Yolov3-t-esd-I row"},{"comment":"All ablations are single runs, and many differences are 0.1-0.3 AP, which is within the typical run-to-run noise of object detection experiments. At least three seeds with means and standard deviations should be reported for the key comparisons, especially the SNI and ESD claims, so that the reader can judge whether the improvements are statistically meaningful.","section":"Sec. 4.2, Tables 1-2"},{"comment":"The final SYOLO architecture is never ablated as a whole from a standard baseline. Because SYOLO changes the backbone (E-ELAN/C2f), the neck (SNI, E-ELAN, GSConvE), and the downsampling (ESD) simultaneously, the final COCO result cannot be attributed to the proposed components. A cumulative ablation (baseline -> +SNI -> +ESD -> +GSConvE -> full SYOLO) with matched or explicitly reported complexity is needed to support the claim that the SA solution causes the state-of-the-art result.","section":"Sec. 4.3 and Table 7"}],"minor_comments":[{"comment":"The phrase 'This paper has been accepted byECCV2024 and published on Springer Nature' is a provenance note rather than scientific content; if this is the published version, cite the ECCV proceedings and clarify the relationship between the arXiv preprint and the published version.","section":"Abstract/header"},{"comment":"The citation [37] is given for PPYoloE, but reference [37] is titled '3D-Man' and appears unrelated; please correct the citation.","section":"Table 6"},{"comment":"The terms 'norm models' and 'lightweight models' are used to choose between ESD-I and ESD-II and between GSConvE-I and GSConvE-II, but they are never defined; give a concrete criterion such as a parameter count or FLOPs threshold.","section":"Sec. 3.3"},{"comment":"The description of training hyperparameters as 'the default set of hyperparameters' is insufficient for reproducibility; please report epochs, batch size, optimizer, learning rate schedule, and augmentation settings for the ablations.","section":"Sec. 4.1"},{"comment":"The comparison between confidence thresholds 0.001 and 0.25 is presented as a 'realistic assessment,' but AP is conventionally computed over a precision-recall curve that includes low-confidence detections; please state the exact evaluation protocol, including whether NMS settings were identical, so readers can interpret the reported AP drop.","section":"Table 3 and Sec. 4.2"},{"comment":"The caption contains a typo: 'Comparasion' should be 'Comparison'.","section":"Table 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be the arXiv version of an ECCV 2024 accepted paper; the editor may wish to verify that the submitted version matches the published camera-ready and that the 'state-of-the-art' claims are consistent with the published record. The main scientific risk is that the SNI mechanism is reparameterizable in the tested necks, so the reported gains may be optimization effects rather than evidence for the misalignment story; the authors should be asked to add the control experiments described in the major comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked about arXiv:2505.12820. I read it. The paper is a useful empirical survey of neck designs for YOLO-style detectors, but the central mechanistic claim does not survive close examination.\n\nWhat is genuinely new: the IHP (an unfused pyramid with bottleneck convs, basically SSD with extra filtering), SNI (a scalar downweighting of nearest-neighbor upsampled features), ESD (a parallel conv+pool downsampling block), and GSConvE variants. The ablation is broad—six YOLO baselines—and the SNI gains are consistent across them. That is real work, and the fact that a one-line change helps at all is worth knowing. ESD and GSConvE also show gains, and the paper is reasonably transparent about the parameter/FLOP changes in the ablation tables.\n\nThe soft spots are real. The stress-test is correct: alpha in Eq. (1) is a spatially uniform scalar, so SNI cannot change which locations are aligned or misaligned. In the concat-and-conv necks used here, scaling one input branch by a constant is equivalent to scaling the corresponding input weights of the first conv; BatchNorm absorbs the rest. The representable function class is unchanged, so the Sec. 3.2 claim that SNI \"confirms the presence of features misalignment\" does not follow. A learned-scalar baseline or an alpha sensitivity study would have at least shown the effect is special to the chosen constant.\n\nThe other issues are also real, if less fatal. The headline comparisons are not matched in capacity: SYOLO has 13.5M more parameters than YOLOv8, and SYolo-s has 3.4M more than YOLOv8-n. That alone could explain part of the gain. All numbers are single runs with no error bars, so the smaller gains (0.2-0.5 AP) are not clearly distinguishable from noise. YOLOv3-t-ESD-I gains an unexplained 11.2 AP with only +3.1M parameters; that outlier is far outside every other result and suggests something other than a clean module effect. And the final SYOLO is not ablated as a whole, so the contributions of IHP/SNI/ESD/GSConvE to the final SOTA number are not established.\n\nOn citation practice: the paper cites the standard FPN line and its own GSConv, which is normal. The abstract's statement that it was \"accepted by ECCV2024 and published on Springer Nature\" is odd for a 2025 arXiv posting; if that is true, this may be a duplicate publication, and it should be verified.\n\nWho gets value: practitioners building real-time detectors will find the SNI trick cheap to try and the IHP results worth knowing. The paper will not reorganize the field, but it is a competent empirical contribution. I would send it to a serious referee, asking for matched-parameter comparisons, error bars, and a learned-scalar ablation to test whether SNI does anything beyond changing the initial scale of one branch.","headline":"A broad empirical study of FPN alternatives that overreaches on mechanism: the SNI scalar is likely a reparameterization of the conv-BN, so the 'feature misalignment' story is unsupported, but the module-level gains may still be useful.","tokens_in":15035,"tokens_out":4260,"would_cite":false,"duration_ms":45487,"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":"FPN-style feature fusion is misaligned; a scalar reweighting fixes it.","keywords":["object detection","feature pyramid","feature misalignment","soft nearest neighbor interpolation","downsampling","lightweight convolution","real-time detection","YOLO"],"falsifier":"Train a fixed detector (for example YOLOv5-n) with and without SNI while holding parameter counts equal, and sweep $\\alpha$ from 0 to 1: if accuracy does not peak near $\\alpha = \\frac{\\text{Resolution}_X}{\\text{Resolution}_Y}$, or if the gain disappears at $\\alpha = 1$, the misalignment explanation fails. Applying SNI to a neck that fuses spatially aligned features (for instance after a learned offset or deconvolution) would further test whether the improvement is specific to misaligned fusion.","tokens_in":14004,"feed_emoji":"🎯","tokens_out":8439,"duration_ms":77681,"temperature":0.7,"pith_summary":"This paper argues that the standard practice of fusing feature maps from different pyramid levels point-to-point in detection necks is fundamentally flawed: low-level features emphasize small and local objects while high-level features emphasize large and global ones, so merging them directly injects noise. To validate this, it builds an unfused independent hierarchy pyramid (IHP) neck that beats fused baselines on coupled-head detectors, then introduces soft nearest neighbor interpolation (SNI), a single scalar multiplication on upsampled high-level features that shrinks their contribution as the resolution gap grows. Combined with a new downsampling module (ESD) and a lightweight convolution (GSConvE), these form the secondary features alignment (SA) solution tested on six YOLO variants. The claim is that this nearly cost-free modification improves accuracy across all tested models and achieves state-of-the-art results on Pascal VOC and MS COCO, suggesting that elaborate fusion architectures may be doing more harm than good.","feed_headline":"FPN-style feature fusion is misaligned; a scalar reweighting fixes it","feed_subtitle":"Soft-scaled upsampling plus new downsampling and lightweight convolutions lifts accuracy on VOC and COCO","key_machinery":"The load-bearing mechanism is a pair of scale-dependent reweighting operations at the two places where feature resolution changes. SNI (soft nearest neighbor interpolation) replaces plain nearest-neighbor upsampling with $Y = \\alpha \\cdot f(X)$, $\\alpha = \\frac{\\text{Resolution}_X}{\\text{Resolution}_Y}$, a scalar that weakens high-level features as the zoom factor grows while keeping nearest-neighbor speed. ESD (features adaptive selection in extended spatial windows) replaces stride-2 downsampling with three parallel branches — a $3\\times3$ convolution, a $4\\times4$ max-pooling, and a $4\\times4$ average-pooling — merged by addition (ESD-I) or learnable linear fusion (ESD-II) to preserve spatial detail. GSConvE simplifies GSConv into variants with large-kernel depthwise branches ($9\\times9$, $13\\times13$, $17\\times17$) for cheap large receptive fields, and IHP (independent hierarchy pyramid) is the diagnostic neck with no fusion, whose behavior isolates the effect of misalignment.","core_discovery":"The central discovery is that feature misalignment in FPN-like necks is systematic, and that a fixed scalar reweighting can mitigate it. For every upsampled feature map $X$ fused into a higher-resolution map $Y$, the paper defines SNI as $Y = \\alpha \\cdot f(X)$ with $\\alpha = \\frac{\\text{Resolution}_X}{\\text{Resolution}_Y}$, so high-level features are downweighted in proportion to how much they are enlarged. Ablations on VOC 07+12 show SNI raising accuracy for every one of the six YOLO baselines, with gains up to +3.3 AP for YOLOv7-tiny and +2.0 AP for YOLOv6-n at no added parameters or FLOPs, and the full SA solution reaches the best accuracy among real-time detectors on both VOC and COCO. The paper also shows that dropping fusion entirely (IHP) helps coupled-head detectors but hurts decoupled-head ones, which it explains by decoupled heads already separating classification and localization features.","pith_inferences":["If SNI acts primarily as a size-dependent gradient regularizer rather than an alignment device, it should also change results when fusing already-aligned feature maps; testing that case would separate the two explanations (the paper does not run this control).","A sensitivity sweep over $\\alpha$ near $\\frac{\\text{Resolution}_X}{\\text{Resolution}_Y}$ could show whether the exact ratio is optimal or merely a good default; a learned per-layer scalar that matches or beats it would suggest the mechanism is adaptive scaling, not the specific formula.","The same misalignment argument should apply to semantic and panoptic segmentation, where boundary regions between object scales are denser, so the SA modules may transfer to those tasks, though the paper only evaluates detection benchmarks."],"forward_implications":["Detectors already deployed with FPN-style necks can gain accuracy by simply swapping the nearest-neighbor upsampling for SNI and keeping everything else fixed.","The IHP result implies that feature fusion is not an obligatory component for coupled-head detectors, so future architecture search can treat fusion as optional.","The ESD result points to downsampling as a recoverable source of spatial information loss, worth revisiting in backbone design beyond the neck.","Consistent gains across six different baselines suggest the misalignment problem is architectural, not specific to one training recipe or dataset."],"supporting_citations":[{"why":"introduces the feature pyramid fusion paradigm that the paper argues is misaligned","marker":"[21]"},{"why":"defines PANet, the multi-level fusion neck used as baseline in YOLO models","marker":"[24]"},{"why":"presents BiFPN, an alternative learnable fusion the paper contrasts with its fixed reweighting","marker":"[34]"},{"why":"defines YOLOv3, one of the six baselines modified to test SNI, ESD and GSConvE","marker":"[31]"},{"why":"defines YOLOv8, a decoupled-head baseline that the IHP hurts and SNI helps","marker":"[15]"},{"why":"defines YOLOv6, where SNI alone adds 2.0 AP in the ablation","marker":"[16]"},{"why":"introduces GSConv, the lightweight convolution that GSConvE extends","marker":"[18]"}],"fun_headline_variants":["Scalar reweighting fixes FPN feature misalignment","Resolution-based scaling repairs feature fusion in detectors","One scalar to fix feature pyramid misalignment","FPN fusion misalignment cured by a simple weight","Downweight upsampled features to align FPN neck"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy gains are attributed to the proposed neck components even though some compared models differ in parameter count and FLOPs, and the SNI downweighting factor $\\alpha$ is fixed by formula without a sensitivity study.","fun_headline_variants_meta":{"raw":{"variants":["Scalar reweighting fixes FPN feature misalignment","Resolution-based scaling repairs feature fusion in detectors","One scalar to fix feature pyramid misalignment","FPN fusion misalignment cured by a simple weight","Downweight upsampled features to align FPN neck"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000553,"raw_usage":{"total_tokens":2642,"prompt_tokens":958,"completion_tokens":1684,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":1610}},"tokens_in":574,"tokens_out":1684,"duration_ms":13445,"temperature":1.0,"reasoning_tokens":1610,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:26:04.616683+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a fixed detector (for example YOLOv5-n) with and without SNI while holding parameter counts equal, and sweep $\\alpha$ from 0 to 1: if accuracy does not peak near $\\alpha = \\frac{\\text{Resolution}_X}{\\text{Resolution}_Y}$, or if the gain disappears at $\\alpha = 1$, the misalignment explanation fails. Applying SNI to a neck that fuses spatially aligned features (for instance after a learned offset or deconvolution) would further test whether the improvement is specific to misaligned fusion.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"defines PANet, the multi-level fusion neck used as baseline in YOLO models"},{"cited_title":"https : / / github","cited_arxiv_id":null,"evidence_quote":"defines YOLOv8, a decoupled-head baseline that the IHP hurts and SNI helps"},{"cited_title":"Journal of Real-Time Image Processing 21(3), 62 (2024)","cited_arxiv_id":null,"evidence_quote":"introduces GSConv, the lightweight convolution that GSConvE extends"}],"review_version":1}