{"id":"34fd3050-7be0-4064-9d3f-f0fb63e716dc","arxiv_id":"2412.03200","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Fab-ME modifies YOLOv8s with a VMamba-based state-space module in the neck and an enhanced channel attention module, reporting 59.4 percent mAP@0.5 on the Tianchi fabric defect dataset versus a 57.4 percent baseline.","lead":"This paper adds two neural network modules to the YOLOv8s object detector and reports a 3.5 percent relative improvement in mean average precision on a 20-class fabric defect benchmark. The work is a modest, incremental engineering contribution that could matter for automated textile quality control if the result generalizes beyond the validation set used.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 3.5% relative gain rests on a single unseeded comparison and a validation set that discards all non-defective tiles; either effect-size noise or the tiling protocol could explain the result, so the improvement is not yet established.","rationale":"The reader's weakest assumption is the baseline reproduction and the validation split. I agree these matter, but the more decisive issue is that a 2.0 mAP-point difference from a single run is not evidence of a real improvement unless repeated-run variance is shown to be smaller. The tiling protocol is a separate validity threat: discarding non-defective tiles removes the false-positive channel, so any method's mAP on this split overstates inspection performance; whether the relative gain survives on full images is unknown. Since neither code, error bars, nor FPS measurements are provided, the central claim is conditional on future release and measurement. This reinforces rather than changes the reader's CONDITIONAL verdict.","tokens_in":8343,"tokens_out":8957,"duration_ms":79522,"concrete_test":"Release the exact config, preprocessing/tiling script, and weights for both Baseline and Fab-ME; then run at least 5 independent training seeds on the same 4:1 split and report mean plus or minus standard deviation of mAP@0.5 together with a 95% confidence interval for the paired difference. Separately evaluate both final models on the original full validation images without tile discarding, merging tile predictions with NMS, and report the same statistics. If the confidence interval of the difference includes zero, or the gain is not reproduced on full-image validation, the claimed 3.5% improvement is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table II and Fig. 3(c) reduce the central claim to one comparison: Baseline 57.4 vs. Fab-ME 59.4 mAP@0.5, a 2.0 absolute / 3.5 relative point gain. No standard deviation, number of seeds, or confidence interval is reported, and the code and trained weights are not released. On YOLO-scale object detectors trained on custom splits, run-to-run variation of this size is common, so the reported gain may be within noise. The evaluation protocol compounds this: Section III-A states that images are segmented into 640x640 sub-images and 'discarding non-defective sub-images' before forming the 3,360 validation images. This removes all true-negative tiles, so mAP is computed only on tiles containing at least one defect. False positives on defect-free fabric do not enter the metric, and tiling also changes effective object scale and context. Finally, the 'real-time' claim in the abstract and Section III-E is asserted without any FPS or latency measurement. Thus the observable support for the central claim is a single, possibly noisy, protocol-specific number.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Fab-ME, a modified YOLOv8s detector for fabric defect detection on the Tianchi dataset. Two architectural changes are introduced: a C2F-VMamba module that replaces one C2F block in the neck by integrating Visual State-Space (VSS) blocks, and an Enhanced Multi-scale Channel Attention (EMCA) module placed after the SPPF block. The paper reports 59.4% mAP@0.5 versus a 57.4% YOLOv8s baseline, i.e., a 3.5% relative improvement, and claims real-time processing. Ablations over insertion positions, the EMCA module, and transfer to YOLOv5s/YOLOv6s are presented.","tokens_in":8578,"tokens_out":4923,"duration_ms":43873,"significance":"If the reported gain is reproducible, this is a modest but potentially useful incremental contribution to textile inspection: a low-parameter YOLO variant with state-space blocks and channel attention. The paper is transparent about the dataset class distribution (Table I) and reports parameter counts. However, the central claim currently rests on a single unseeded validation split, and the evaluation protocol discards all non-defective sub-images, so the 3.5% improvement is not yet established. In addition, the real-time claim is unsupported by any speed measurement. With these gaps addressed, the contribution would be of interest to the defect-detection community.","major_comments":[{"comment":"The validation protocol discards non-defective 640x640 sub-images before computing mAP@0.5. As a result, the validation set contains only tiles with at least one defect, so false positives on defect-free fabric are never penalized in the reported metric. Since the intended application is continuous fabric inspection, this protocol may overstate practical detection quality. Please report results on the full original test images, or at least on a keep-all-tiles split, as a sensitivity check.","section":"Section III-A"},{"comment":"The central claim rests on a single comparison: Baseline 57.4 versus Fab-ME 59.4 mAP@0.5, with no standard deviation, number of seeds, or confidence interval. For modern object detectors, run-to-run variation of this magnitude is common, and the insertion positions were selected on the same validation set, which adds a selection effect. Please run multiple seeds and report mean and standard deviation, or provide a bootstrap confidence interval at minimum. Releasing code and trained weights would also make the result independently verifiable.","section":"Table II and Fig. 3"},{"comment":"The abstract and Section III-E.4 claim real-time processing, but no frames-per-second or latency measurement is reported anywhere in the paper. Since C2F-VMamba adds a 2D selective-scan operation and EMCA adds a convolution-based attention module, the speed claim needs direct support. Please measure and report inference time on the stated RTX 4090 hardware, including preprocessing time if the tiling protocol is part of deployment.","section":"Abstract and Section III-E.4"},{"comment":"The EMCA module is described as 'multi-scale', but the displayed computation is a single fixed-size 1D convolution applied to the sum of two global pooling descriptors: a = sigma(Conv1D(GAP(F) + GMP(F), k)). There are no multiple kernel sizes, dilation rates, or parallel branches. Either clarify what makes this attention multi-scale, or replace the 'multi-scale' characterization with a more accurate description such as 'dual-pooling channel attention'.","section":"Section II-C"},{"comment":"The text states that replacing the third C2F block gives a 'performance improvement of 2.5%'. If the values in Fig. 3(a) are read as mAP percentages, the baseline is 57.4 and the best single-position replacement is 58.9, which is 1.5 absolute points or about 2.6% relative, not 2.5%. The paper should state explicitly whether all reported percentages are absolute or relative and should reconcile the text with the figure labels.","section":"Section III-E.1 and Fig. 3(a)"}],"minor_comments":[{"comment":"The text first says EMCA is integrated 'after the SPPF block within the YOLOv8 backbone', but the following sentence says the light blue box denoting EMCA is in the 'feature fusion module neck'. Fig. 1 appears to show EMCA after a concatenation in the neck. Please clarify the actual insertion location.","section":"Section II-A"},{"comment":"Section III-F says 'The comparison diagram in Fig. 2 presents the predictions', but Fig. 2 is a sample display of original defect images and Fig. 5 is the visualization figure. The reference should be corrected to Fig. 5.","section":"Section III-F"},{"comment":"The introduction cites [23] for Vision Mamba and [25] for VMamba, which are distinct models with the same first-author group. The text should make clear which specific architecture is used for the VSS blocks in C2F-VMamba.","section":"Section I"},{"comment":"Fig. 4 reports results on YOLOv5s and YOLOv6s, but the text in Section III-E does not explain the training setup or the 'Original' versus 'Ours' comparison for those backbones. Adding a sentence on the hyperparameters and evaluation split used for Fig. 4 would improve reproducibility.","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal and the empirical claim is plausible, but the missing error bars, the non-defective-tile discard protocol, and the unsupported real-time assertion all need to be fixed before the central claim can be accepted. I do not see evidence of misconduct; the issues are experimental rigor and internal consistency."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a cleanly executed but under-evidenced engineering paper. The claimed 3.5% relative mAP gain over YOLOv8s on the Tianchi fabric dataset is plausible but not established; the paper lacks repeated runs, error bars, and a clean evaluation protocol.\n\nWhat's good: the combination of VMamba's VSS blocks into the C2F neck module and the ECA variant with dual pooling (EMCA) is a sensible, well-specified integration. The ablation in Fig. 3 varies the insertion point of C2F-VMamba across the four neck positions, which is the right way to justify a design choice, and the result that the third position works best is informative. The modules are described precisely enough to reimplement, and the authors reused existing state-space and attention components rather than inventing new ones. That counts for something.\n\nThe soft spots are real. There is exactly one validation run per configuration. On a custom split of a fabric dataset, a 2-point mAP difference falls inside typical run-to-run noise for YOLO-scale detectors, so we cannot rule out that the gain is selection noise. The evaluation protocol is worse: Section III-A says non-defective sub-images are discarded after tiling. That means the validation set has no true negatives, so false positives on undefected fabric never count against the model. This biases mAP upward and breaks comparability with the results quoted from other papers in Table II, which used their own, likely different, evaluation protocols. The 'real-time' claim is asserted without any FPS or latency number. Minor: the ablation descriptions in Sec. III-E1 say a 2.5% improvement for the best C2F position, while Fig. 3(a) shows 58.9 vs 57.4 (about 2.6% relative), and the inconsistent caption style makes the figure hard to read.\n\nNone of this means the method is bad or the idea is wrong. VMamba in the neck is a reasonable thing to try for long, narrow fabric defects, and the paper gives a useful datapoint. But as it stands the central claim is not independently verifiable.\n\nI would send this to peer review because the question is legitimate and the ablation is informative; a referee should require code, multiple seeds, and an evaluation that keeps the non-defective tiles. Then it could become a solid applied contribution.","headline":"Plausible engineering gain, unproven: single-seed numbers and a defect-only validation set leave the 3.5% mAP improvement inside noise.","tokens_in":9130,"tokens_out":3001,"would_cite":false,"duration_ms":28173,"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 proposes Fab-ME, a YOLOv8s variant with C2F-VMamba and EMCA modules, reaching 59.4% mAP@0.5 on the Tianchi fabric defect dataset, a 3.5% relative gain over the YOLOv8s baseline.","keywords":["fabric defect detection","YOLOv8s","vision state-space model","VMamba","channel attention","real-time object detection","multi-scale feature extraction","Tianchi fabric dataset"],"falsifier":"Run both YOLOv8s and Fab-ME from the same codebase on the identical tiled Tianchi train/validation split with the same hyperparameters and multiple random seeds; if the average mAP@0.5 gap is not close to 2.0 absolute points, or if confidence intervals overlap, the central claim fails.","tokens_in":8134,"feed_emoji":"🧵","tokens_out":8119,"duration_ms":64811,"temperature":0.7,"pith_summary":"This paper is trying to show that a lightweight modification of YOLOv8s can detect fabric defects more accurately without sacrificing speed. The proposed Fab-ME builds global context into the neck with C2F-VMamba, a module that inserts vision state-space blocks into YOLOv8s' C2F block, and improves small-target sensitivity with EMCA, a multi-scale channel attention placed after SPPF. On the Tianchi 20-class fabric defect dataset, Fab-ME reaches 59.4% mAP@0.5, beating the 57.4% YOLOv8s baseline by 3.5% relative, while keeping about 11 million parameters and real-time speed. If correct, it offers a practical recipe for industrial fabric inspection, where defects are long, narrow, and varied.","feed_headline":"Fabric defect detector hits 59.4 mAP with state-space and attention","feed_subtitle":"Adds VMamba and multi-scale attention to YOLOv8s, beating its baseline by 3.5 percent at real-time speed.","key_machinery":"The machinery is the C2F-VMamba module and the EMCA module. C2F-VMamba takes the YOLOv8s C2F, the cross-stage partial bottleneck with two convolutions, and routes one split through stacked VSS blocks; each VSS block splits again, sends one branch through a depthwise convolution and the other through a 2D selective scan (SS2D), then merges them. EMCA computes channel weights from the sum of global average pooling and global max pooling passed through a 1D convolution with kernel size 5 and a sigmoid, then multiplies each channel. The first supplies a global receptive field at linear complexity, and the second recalibrates channels with multi-scale descriptors.","core_discovery":"Fab-ME is a YOLOv8s detector in which the third C2F block in the neck is replaced by C2F-VMamba, a C2F module whose split path runs through stacked visual state-space (VSS) blocks built on 2D selective scanning (SS2D), and in which the SPPF output of the backbone is followed by EMCA, a channel attention that sums global average and global max pooling before a 1D convolution with kernel size 5 and a sigmoid. On the Tianchi 20-class fabric defect dataset this reaches 59.4% mAP@0.5, compared with a 57.4% YOLOv8s baseline, a 3.5% relative improvement, while parameters stay near 11.0 million. The paper credits the VSS blocks for global context and detail capture and EMCA for small-target sensitivity, with the ablation showing that replacing the third neck C2F contributes the larger share of the gain.","pith_inferences":["Beyond the paper's evidence: the modules are not fabric-specific, so the same recipe is plausible for steel, wood, or ceramic surface-defect detection, but the paper does not test this.","Beyond the paper's evidence: because non-defective 640x640 tiles are discarded before training and validation, the reported numbers describe a defect-dense distribution; on untrimmed production images the gain could shrink.","Beyond the paper's evidence: rerunning the baseline and Fab-ME with multiple seeds and reporting variance would establish whether the 2.0 absolute-point gap is stable or within run-to-run noise."],"forward_implications":["On the Tianchi benchmark, Fab-ME reaches 59.4% mAP@0.5, ahead of the 57.4% YOLOv8s baseline and every listed published method in the comparison table.","The ablation attributes the largest single gain to replacing the third C2F in the neck with C2F-VMamba, with EMCA adding a further increment.","The same two modules also improve YOLOv5s and YOLOv6s in the authors' experiments, so the recipe transfers across YOLO versions.","Parameter count stays near 11.0 million, so the added modules do not obviously cost real-time throughput."],"supporting_citations":[{"why":"provides the VSS blocks and SS2D operation that C2F-VMamba embeds in the neck","marker":"[25]"},{"why":"defines YOLOv8s and its C2F architecture that Fab-ME modifies","marker":"[26]"},{"why":"supplies the ECA baseline that EMCA extends with average-plus-max pooling and a 1D convolution","marker":"[28]"},{"why":"names the SPPF block after which the paper inserts EMCA","marker":"[27]"},{"why":"motivates bidirectional state-space visual representation for the global-context design","marker":"[23]"}],"fun_headline_variants":["State-space + attention lift fabric defect mAP to 59.4","YOLOv8s upgrade: VMamba blocks and EMCA boost defect mAP by 3.5%","Fab-ME: VMamba and attention beat baseline on fabric defects","Defect detection: C2F-VMamba and EMCA push mAP to 59.4","Fabric defects spotted better: 59.4 mAP with VMamba neck"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gain assumes the 57.4% YOLOv8s baseline was trained and validated under exactly the same tiled, non-defective-discarded protocol as Fab-ME; if that baseline reproduction is off, the claimed 3.5 percent improvement is not established.","fun_headline_variants_meta":{"raw":{"variants":["State-space + attention lift fabric defect mAP to 59.4","YOLOv8s upgrade: VMamba blocks and EMCA boost defect mAP by 3.5%","Fab-ME: VMamba and attention beat baseline on fabric defects","Defect detection: C2F-VMamba and EMCA push mAP to 59.4","Fabric defects spotted better: 59.4 mAP with VMamba neck"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000685,"raw_usage":{"total_tokens":3117,"prompt_tokens":962,"completion_tokens":2155,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":2042}},"tokens_in":578,"tokens_out":2155,"duration_ms":13180,"temperature":1.0,"reasoning_tokens":2042,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:40:20.345849+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both YOLOv8s and Fab-ME from the same codebase on the identical tiled Tianchi train/validation split with the same hyperparameters and multiple random seeds; if the average mAP@0.5 gap is not close to 2.0 absolute points, or if confidence intervals overlap, the central claim fails.","supporting_citations":[{"cited_title":"A review on yolov8 and its advancements,","cited_arxiv_id":null,"evidence_quote":"defines YOLOv8s and its C2F architecture that Fab-ME modifies"},{"cited_title":"Tph-yolov5: Improved yolov5 based on transformer prediction head for object detection on drone-captured scenarios,","cited_arxiv_id":null,"evidence_quote":"names the SPPF block after which the paper inserts EMCA"}],"review_version":1}