{"id":"81fa4eb4-31f7-44dd-9d0d-069ac76d158c","arxiv_id":"2412.14790","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Pruning YOLOv11's detection heads yields six size-specialized variants that match original accuracy on targeted datasets while cutting model size, FLOPs, and inference time.","lead":"The authors modify YOLOv11 by pruning its detection heads to create six model versions, each specialized for small, medium, or large objects. The result is a set of smaller, faster models that hold accuracy close to the original on six public datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central accuracy-preservation claim rests on an untested separability assumption: no per-size AP is reported, and aggregate mAP on size-skewed datasets cannot show that pruning leaves target-size accuracy unchanged.","rationale":"The reader's conditionality is well-founded. The efficiency metrics (model size, GFLOPs, inference time) follow almost mechanically from removing blocks, so I see no significant concern there. The load-bearing part is accuracy preservation. The paper's datasets are selected specifically because they are dominated by the target size, which makes aggregate mAP an indirect and weak test of the pruning rationale. Moreover, the text contains a concrete internal inconsistency about which feature-map levels detect small versus large objects (Section 3), which makes it important to verify the actual implemented mapping rather than take the prose at face value. The single-seed protocol and the absence of error bars further mean that the reported gains or losses around 1-2% cannot be distinguished from training noise. None of this proves the central claim false; it means the current evidence is insufficient to accept it as a definitive result. A per-size AP evaluation would settle whether the pruning assumption holds.","tokens_in":11092,"tokens_out":11189,"duration_ms":76267,"concrete_test":"Re-run the Table 4 experiments and report AP@50 separately for small, medium, and large object subsets, using the area thresholds in Table 1, over at least three training seeds. The separability assumption is supported if, on its target subset, the pruned model's AP is within the seed variance of the full model's AP on that same subset; it is refuted if target-subset AP drops by more than 2 points.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's efficiency improvements are structural and credible, but the central claim that the pruned models 'maintain accuracy' of YOLOv11 rests on an untested separability assumption. In standard YOLOv11, the three detection heads sit after a PANet neck that fuses features across scales; the P3 (small-object) head receives input that is a concatenation of backbone P3 and upsampled P4 features, so removing the medium/large branches (Section 3.1, blocks b17-b22) can alter the retained head's input, not just discard irrelevant computation. The paper's only evidence is aggregate mAP@50 on datasets heavily skewed toward the target size (Table 2: WeedCrop 81% small, Underwater Pipes 96% large, Brain Tumor 79% large). Aggregate mAP on such datasets is dominated by the target size, so it cannot reveal whether target-size AP is preserved or whether the model now misses non-target sizes. No per-size AP or block-removal ablation is reported. The inconsistency in Section 3's feature-map semantics (20x20 described as detecting small objects, the opposite of standard YOLO) heightens the risk that the pruning rationale is based on a wrong mapping. With single-seed runs (Seed 0, Table 3), the sub-2% differences in Table 4 cannot be distinguished from noise. Thus the claim that pruned variants can replace the full model for fixed object-size distributions is not directly supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes six size-specific pruned variants of YOLOv11n (small, medium, large, sm, ml, sl) by removing blocks in the neck and head that are claimed to serve non-target object sizes, together with a small program that classifies objects in a dataset by size and recommends one of the variants. The variants are evaluated on six datasets and compared with YOLOv11, YOLOv10, and YOLOv8 in terms of recall, precision, mAP@50, model size, GFLOPs, inference time, and power consumption. The central claim is that the pruned models maintain the accuracy of the full YOLOv11 while substantially reducing resource use.","tokens_in":11328,"tokens_out":6620,"duration_ms":43449,"significance":"If the accuracy-preservation claim were established, the proposed variants would be a useful practical contribution for applications with a fixed, known object-size distribution, and the public code and weights would support reproducibility. The efficiency gains are structurally credible: removing detection heads and neck branches necessarily reduces parameter count and computation. However, the current evidence does not support the central accuracy claim because only aggregate mAP on size-skewed datasets is reported, no per-size AP is provided, the experiments use a single seed, and the pruning rationale contains an internal inconsistency about which feature levels detect which object sizes. The paper therefore needs a substantial experimental revision before the central claim can be accepted.","major_comments":[{"comment":"The claim that the pruned variants 'maintain accuracy' rests on an untested separability assumption: that the three detection heads and their feeding blocks are functionally separable by object size, so that removing all blocks for non-target sizes leaves target-size accuracy unchanged. No per-size average precision is reported. Table 4 gives only aggregate mAP@50 on datasets that are heavily skewed toward the target size (Table 2: WeedCrop has 15,237/18,693 small objects; Underwater Pipes has 11,683/12,238 large; Brain Tumor has 16,985/21,526 large). Aggregate mAP on such datasets is dominated by the dominant size and cannot reveal whether target-size AP is preserved or whether the model now misses non-target sizes. Please report per-size AP for every dataset/model, or evaluate on a size-balanced benchmark, and provide an ablation that removes one block group at a time.","section":"§3.1–3.6 and Table 4"},{"comment":"The feature-map semantics used to justify pruning are internally inconsistent and opposite to standard YOLO. Section 2 states that 'the smallest feature sets, such as 20x20, focus on detecting small objects,' whereas in standard YOLO the 20x20 / stride-32 head detects large objects and the 80x80 / stride-8 head detects small objects. Section 3.1 then calls the 'first detection head' the small-object head. Because the choice of which blocks to remove (b14–b16 versus b17–b22) depends on which head corresponds to which object size, an incorrect mapping would remove the wrong heads and invalidate the architectural rationale. Please state explicitly, for each of the three heads, the feature level (P3/P4/P5), stride, and object-size assignment, and reconcile this with the pruning descriptions.","section":"§2 and §3"},{"comment":"The computational-cost values are reported with incorrect units. The abstract says 'the required GFLOPs were reduced from 6.3MB (YOLOv11), 5.7MB (YOLOv10) and 8.1MB (YOLOv8) to just 3.8MB for the large model,' but these numbers appear to be model sizes in megabytes, not GFLOPs. Section 5.2 and Figure 11 also use 'GFLOPS' interchangeably with 'GFLOPs' and describe it as 'giga floating-point operations per second.' Please report the actual GFLOPs per inference for each model in the proper units, and present model size in MB separately, as in Figure 10.","section":"Abstract and §5.2, Figure 11"},{"comment":"The accuracy comparison is based on a single seed (Seed 0, Table 3) with no error bars or repeated runs, and the maximum differences in Table 4 are generally under 2%, with some entries equal (e.g., Face Detection recall is 93.10 for both YOLOv11 and YOLOv11-sl). These differences cannot be distinguished from training noise. Please run at least three seeds per condition and report mean and standard deviation, or otherwise provide a significance test, before claiming that the modified models 'maintain' or 'outperform' the original YOLOv11.","section":"§5.1, Tables 3 and 4"},{"comment":"There is a factual inconsistency about which model was used for the Brain Tumor dataset. Section 4.2 says 'the YOLOv11-sm modified version is used for the evaluation process of this dataset,' but Table 2 and Table 4 both list YOLOv11-ml for Brain Tumor. Relatedly, Section 5.1 states that 'the original YOLOv11 outperforms the modified versions (YOLOv11-medium and YOLOv11-sm),' while Table 4 shows YOLOv11-sm achieves higher mAP@50 than YOLOv11 on Aerial Airport (93.20 vs. 92.93); the text should refer to YOLOv11-ml. Please correct these inconsistencies, since the model-selection procedure is part of the contribution.","section":"§4.2, Tables 2 and 4, §5.1"}],"minor_comments":[{"comment":"There are several typos and inconsistent terms: 'Limition' in Section 7, 'blcok' in §3.5, 'V alue' in Table 3, 'YOLO10' versus 'YOLOv10' in §5.1, and 'measures the abbreviation measures' in §5.2. A careful proofreading pass is needed.","section":"Global"},{"comment":"The measurement protocol for inference time and power consumption is not described. Please specify the hardware, software versions, batch size used for timing, and how 'Watts per epoch' was measured; this is needed for the resource-efficiency claims to be reproducible.","section":"§5.2 and Figure 13"},{"comment":"The Aerial Airport dataset is cited as [7,12] in the text but as [12] in Table 2, and the reference numbering should be aligned across the body, tables, and figures.","section":"References"},{"comment":"The object-size classifier is described only in prose; providing the exact decision rule and an example output would make the model-selection step more transparent and reproducible.","section":"§4.1"},{"comment":"Figure 11 is labeled 'Computational Performance (GFLOPS)' but the text discusses 'GFLOPs'; please use one term consistently and define it as either total operations per inference or operations per second.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is an engineering contribution with limited novelty: the pruning is straightforward and the main open question is whether the size-based separability assumption holds. The experimental evidence is currently too weak to support the central claim, but the issues are fixable with additional experiments (per-size AP, multi-seed runs, corrected units, and a corrected feature-level mapping). The self-citation in reference [25] appears unrelated to the technical content and could be removed. Given the journal context, the paper may be more suitable as a short application note if the accuracy-preservation claim is adequately substantiated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the efficiency numbers are real, the accuracy-preservation claim is not yet supported. The paper proposes six pruned YOLOv11n variants, each keeping only the detection heads (and feeding blocks) for a specified object-size range, plus a small script that counts bounding-box areas to choose a variant. That is a reasonable engineering recipe, executed across six datasets and compared against YOLOv10 and YOLOv8. Model size, GFLOPs, and inference time all drop as expected from structural pruning, so the resource-utilization story is credible.\n\nThe soft spot is the central claim that the pruned models 'maintain accuracy.' Table 4 reports aggregate mAP@50 on datasets that are overwhelmingly skewed toward each model's target size—WeedCrop is 81% small, Underwater Pipes is 96% large, Brain Tumor is 95% medium+large. Aggregate mAP on such data is dominated by the target size, so a model that completely fails on non-target sizes can still show nearly unchanged aggregate mAP. There is no per-size AP anywhere, no block-removal ablation, and no check that the retained heads still receive the same feature inputs. Since the YOLOv11 neck fuses features across scales (the P3 head takes upsampled P4 features, for example), removing medium/large branches can alter the retained head's input rather than just discarding irrelevant computation. The separability assumption is load-bearing and untested.\n\nThe other issues are smaller. Section 3 describes the 20x20 feature map as detecting small objects, which is the opposite of the usual YOLO semantics; this makes the pruning rationale hard to trust, even if the implementation happens to be correct. Runs are single-seed with no error bars, so the sub-2% mAP differences in Table 4 are within noise. The abstract reports GFLOPs in MB. The object-size classifier is just a histogram; it is not a contribution. The citation pattern is thin, but that is not where the paper's problem lies.\n\nNone of this proves the pruning idea is wrong. For fixed object-size domains, dropping irrelevant heads from YOLOv11 is a sensible move, and the efficiency gains are exactly what you would expect. But the evidence for accuracy preservation is missing. The paper needs per-size AP, a feature-flow verification that retained heads still get their original inputs, and multi-seed runs.\n\nWho should read this: practitioners working on edge deployment with known object-size distributions. It deserves a serious referee despite the flaws—the pruning question is real and the experiments are extensive—but the referee should insist on the per-size analysis. I would bring it to a reading group as a cautionary example of aggregate metrics on skewed data.","headline":"A plausible pruning recipe whose efficiency claims are credible, but the 'maintains accuracy' claim is not supported because aggregate mAP on size-skewed data masks the untested separability assumption.","tokens_in":11873,"tokens_out":5312,"would_cite":false,"duration_ms":33847,"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":"YOLOv11 can be pruned by object size into six smaller variants that keep accuracy while using fewer resources.","keywords":["YOLOv11","object detection","network pruning","size-specific detection","model compression","inference time","resource efficiency","object size classification"],"falsifier":"Take a dataset containing objects of all three size classes, train both the full YOLOv11 and one pruned variant (for example YOLOv11-small) on the same data, then compute mAP@50 restricted to small objects only. If the pruned model's small-object mAP falls more than about two points below the full model's, the head-separability assumption is wrong.","tokens_in":1563,"feed_emoji":"🎯","tokens_out":1890,"duration_ms":47840,"temperature":0.7,"pith_summary":"The paper claims that YOLOv11's three detection heads are specialized by object size, and that pruning the heads and feeding blocks for sizes absent from a dataset yields models that match or slightly beat the original's accuracy while using fewer resources. The authors build six variants, corresponding to small, medium, large, small-medium, medium-large, and small-large, plus a small program that counts object areas in a dataset's labels to pick the right variant. On six datasets, the pruned models keep mAP@50 within about two percentage points of the full model, while reducing model size, GFLOPs, inference time, and, in most cases, power consumption. If this holds, object-detection deployments with a fixed mix of object sizes can replace the full network with a smaller tailored one.","feed_headline":"Six size-tuned YOLOv11 models match accuracy with far fewer resources","feed_subtitle":"Pruning detection heads for absent object sizes keeps mAP within 2% while cutting model size, GFLOPs, and inference time.","key_machinery":"The load-bearing mechanism is the YOLOv11 head structure: three detection blocks each consuming feature maps at a particular scale (80x80, 40x40, and 20x20), conventionally associated with small, medium, and large objects. The authors label every architectural block from b0 to b22, then for each variant delete the detection head (or heads) for the sizes they want to drop, along with the upstream blocks that feed those heads, reconnecting surviving paths to keep the network a valid feed-forward graph. The object-size classifier program reads label files, bins each bounding-box area into the paper's size categories, and determines which variant matches a dataset's dominant object sizes.","core_discovery":"The central claim is that object-size specialization is built into the YOLOv11 architecture: the three detection heads receive feature maps at different scales (80x80, 40x40, 20x20), so blocks that feed a head for a size range not present in a dataset can be removed without materially hurting detection of the sizes that remain. In experiments across six datasets, each pruned variant's mAP@50 stayed within roughly two percentage points of the full YOLOv11, and four of the six variants actually surpassed it. Model size fell from about 6.3 MB to as low as 3.4 MB, GFLOPs dropped substantially, and inference time decreased in every case.","pith_inferences":["The paper's limitation note implies that for scenes with continuously varying object sizes, such as cars on a road growing and shrinking with distance, the full model remains safer; a natural extension would be a size-aware router that switches among pruned variants per image or per frame rather than per dataset.","The claim that pruning heads for absent sizes leaves target-size accuracy intact could be tested more directly by measuring per-size mAP (small, medium, large) rather than the aggregate mAP@50 the paper reports.","If the head-separability claim transfers across YOLO generations, the same block-removal strategy could be applied to YOLOv8, YOLOv10, or future versions, potentially without retraining from scratch."],"forward_implications":["For deployments where object sizes are roughly fixed, such as aerial imagery, medical cells, or industrial inspection, a pruned variant can replace the full YOLOv11 with no meaningful accuracy loss and a smaller memory footprint.","The same pruning recipe can be applied to the larger YOLOv11 variants (s, m, l, x), since the modifications target a shared block structure.","The object-size classifier makes model selection automatic: given a dataset's labels, it counts instances per size bin and suggests the smallest architecture whose size coverage fits that distribution.","Reduced GFLOPs and faster inference make the pruned models better suited to edge or real-time deployments than the original YOLOv11.","Combined variants (sm, ml, sl) cover bimodal size distributions, extending the approach beyond single-size datasets."],"supporting_citations":[{"why":"Ultralytics' documented size thresholds define the paper's object-size categories (small up to 32x32, medium to 96x96, large above).","marker":"[31]"},{"why":"Alif's analysis of YOLOv11 for vehicle detection supplies the justification that specific feature-map scales correspond to specific object-size detections.","marker":"[4]"},{"why":"Feng et al.'s improved YOLOv8 for small objects in aerial imagery is the other citation used to justify the size-to-head correspondence.","marker":"[10]"},{"why":"Ghosh's YOLOv11 architecture overview provides the block diagram the authors label b0 to b22 for pruning decisions.","marker":"[13]"},{"why":"Khanam and Hussain's overview of YOLOv11 underpins the description of the backbone, neck, and head components that the pruning targets.","marker":"[19]"},{"why":"Ultralytics' official YOLOv11 documentation supplies the baseline model used in the accuracy and efficiency comparisons.","marker":"[32]"}],"fun_headline_variants":["Prune YOLOv11 heads for absent sizes, keep mAP, cut GFLOPs","Size-aware pruning of YOLOv11 trims model and speeds inference","YOLOv11 pruned per dataset's object sizes: lighter, faster, accurate","Removing unused YOLOv11 detection heads saves resources, not accuracy"],"cache_read_input_tokens":13952,"weakest_assumption_plain":"The argument assumes that the three detection heads of YOLOv11 operate independently by object size, so deleting the blocks that feed the heads for sizes absent from a dataset does not remove features that the remaining heads still need; if the heads share features through the neck, accuracy on the target size will drop.","fun_headline_variants_meta":{"raw":{"variants":["Prune YOLOv11 heads for absent sizes, keep mAP, cut GFLOPs","Size-aware pruning of YOLOv11 trims model and speeds inference","YOLOv11 pruned per dataset's object sizes: lighter, faster, accurate","Removing unused YOLOv11 detection heads saves resources, not accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001215,"raw_usage":{"total_tokens":4955,"prompt_tokens":855,"completion_tokens":4100,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":4024}},"tokens_in":471,"tokens_out":4100,"duration_ms":22126,"temperature":1.0,"reasoning_tokens":4024,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:54:19.819889+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset containing objects of all three size classes, train both the full YOLOv11 and one pruned variant (for example YOLOv11-small) on the same data, then compute mAP@50 restricted to small objects only. If the pruned model's small-object mAP falls more than about two points below the full model's, the head-separability assumption is wrong.","supporting_citations":[{"cited_title":"Yolov11 classification of ob- ject sizes","cited_arxiv_id":null,"evidence_quote":"Ultralytics' documented size thresholds define the paper's object-size categories (small up to 32x32, medium to 96x96, large above)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Feng et al.'s improved YOLOv8 for small objects in aerial imagery is the other citation used to justify the size-to-head correspondence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Ghosh's YOLOv11 architecture overview provides the block diagram the authors label b0 to b22 for pruning decisions."},{"cited_title":"Yolov11 documentation, 2024","cited_arxiv_id":null,"evidence_quote":"Ultralytics' official YOLOv11 documentation supplies the baseline model used in the accuracy and efficiency comparisons."}],"review_version":1}