Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper introduces LUSITANOv2, a large-scale fabric defect dataset captured with industrial line-scan cameras directly on an operational textile production line in Portugal, expanding upon the prior LUSITANO dataset. The authors detail six-month on-site data collection workflow, YOLO-format single-class bounding box annotations, comprehensive statistical analysis of dataset visual characteristics (defect count distribution, defect coverage, boundary cropping issues, train/val/test split consistency), and benchmark evaluations on two mainstream detection paradigms: supervised object detectors (YOLOv12, Faster R-CNN, DETR) and unsupervised one-class anomaly detection models (SimpleNet, PatchCore, SuperSimpleNet, etc.). The core claim is that LUSITANOv2 is the largest realistic industrial line-scan textile defect dataset to date, serving as a challenging benchmark to advance practical fabric inspection algorithms.

 

+Strengths

1. Unlike most public textile datasets captured under controlled lab environments with flat scanners/area cameras, the authors deploy a professional Teledyne Dalsa industrial line-scan camera on a running production line with factory LED linear lighting. Raw data naturally contains real interferences including uneven illumination, fabric tension deformation, motion blur, vignetting, and partial border defects, which closely reproduces real factory inspection difficulties and improves dataset practical value.
2. The work provides multi-dimensional quantitative statistics for the dataset, including defect quantity per image, defect pixel coverage distribution, detection difficulty metrics (tiny defects, border-truncated defects, overlapping defects), brightness/contrast balance across train/validation/test partitions, and KS-test verification of split statistical homogeneity. Visual histograms and heatmaps clearly illustrate core dataset challenges (68.06% defects touch image borders), offering transparent baseline statistics for follow-up researchers.
3. The paper evaluates both mainstream supervised object detection pipelines and state-of-the-art unsupervised anomaly detection methods using the open-source Anomalib library, reporting standardized mAP@0.5 and AUROC metrics respectively. Baseline results clearly demonstrate that current mainstream models suffer severe performance degradation on real industrial textile data, proving the dataset’s difficulty and avoiding the saturation problem common to small lab datasets.
4. The dataset adopts a widely accepted YOLO folder structure split into train/val/test subfolders for images and matching bounding box labels, which lowers the threshold for reuse by computer vision researchers focused on industrial defect detection. All captured defects are natural factory defects without artificial synthesis or data augmentation, retaining true real-world defect distribution.

+Weaknesses

1. The most prominent limitation is uniform single-class "defect" labeling for all anomalies (slubs, neps, holes, stains, mispicks, yarn breakages, tension marks, etc.). Existing public datasets including TILDA, AFID and ZJU-Leaper all provide multi-class defect labels to support classification tasks, yet LUSITANOv2 abandons this essential granularity. The authors acknowledge this limitation but provide no viable solution or supplementary fine-grained labels, drastically reducing the dataset’s research scope and innovation value.
2. The paper only expands image volume from 2,875 to 25,114 without innovative collection hardware, annotation schema, or industrial scene expansion. No new textile material types, defect categories, production line conditions, or imaging modalities are introduced. The core dataset characteristics (line-scan camera, single bounding box annotation, single defect class) remain identical to LUSITANO, making the work a simple data scaling exercise rather than an independent novel contribution worthy of standalone journal publication.
3. The comparative table of public datasets only lists basic volume metrics but fails to conduct in-depth qualitative comparison with AITEX and RAW-FABRID (existing line-scan datasets). The authors do not quantitatively analyze how LUSITANOv2 outperforms prior line-scan benchmarks in defect diversity, imaging interference richness, or annotation quality, weakening the core argument of "largest and most comprehensive line-scan textile dataset".
4. Unreasonable input resolution and inconsistent experimental hyperparameter settings. The experiments use contradictory input resolutions (256×256 for general setup description vs. 512×128 actual input for line-scan images) without ablation studies to justify this compression choice. Line-scan fabric images are ultra-long horizontal strips; aggressive downsampling inevitably erases tiny subtle defects, yet the authors never analyze how resolution compression harms model performance, making benchmark results untrustworthy for real industrial deployment reference.
5. The core selling point of LUSITANOv2 is its real industrial domain shift challenge, yet the authors do not conduct cross-domain transfer experiments (training on lab datasets TILDA/ZJU-Leaper and testing on LUSITANOv2, or vice versa). There is no quantitative proof that models trained on lab data suffer severe performance drop on this dataset, which invalidates the central claim that the benchmark addresses industrial domain shift gaps.
6. The dataset does not store corresponding production parameters (fabric material, weaving type, production speed, lighting adjustment records, raw yarn specifications) captured during image acquisition. These metadata are critical for researching causal analysis of defect formation and domain adaptive algorithms, yet the authors discard all contextual factory data, limiting multi-modal and industrial causal research based on this dataset.

 

 

 

 

 

 

 

Author Response

We thank you for your clear and detailed feedback. We have revised the manuscript, in order to address your criticisms, and believe you will be satisfied with the current state. 

We now answer your criticisms, namely the "Weaknesses", point for point:

• The most prominent limitation is uniform single-class "defect" labeling for all anomalies (slubs, neps, holes, stains, mispicks, yarn breakages, tension marks, etc.). Existing public datasets including TILDA, AFID and ZJU-Leaper all provide multi-class defect labels to support classification tasks, yet LUSITANOv2 abandons this essential granularity. The authors acknowledge this limitation but provide no viable solution or supplementary fine-grained labels, drastically reducing the dataset’s research scope and innovation value.

We agree that fine-grained defect labels are valuable for supervised classification and category-specific analysis. We have revised Section 3.1 to clarify that LUSITANOv2 provides class-agnostic defect localization rather than a defect taxonomy. References to slubs, holes, stains, and other defects describe visually observed examples rather than released annotation categories.

The current annotation scheme supports bounding-box localization and, together with the defect-free images, one-class anomaly detection. These tasks are relevant when the objective is to identify and localize departures from acceptable fabric appearance without assigning a predefined defect category.

We acknowledge that this scope does not replace multi-class datasets or resolve the need for supplementary fine-grained annotations. Section 6 explicitly identifies the absence of fine-grained classes, severity labels, and pixel-level masks as limitations and identifies richer annotations as a future extension. We have therefore clarified the supported research tasks without claiming that fine-grained classification is supported by the present release.

• The paper only expands image volume from 2,875 to 25,114 without innovative collection hardware, annotation schema, or industrial scene expansion. No new textile material types, defect categories, production line conditions, or imaging modalities are introduced. The core dataset characteristics (line-scan camera, single bounding box annotation, single defect class) remain identical to LUSITANO, making the work a simple data scaling exercise rather than an independent novel contribution worthy of standalone journal publication.

We appreciate this concern and have clarified the nature of the contribution. LUSITANOv2 extends LUSITANO; the acquisition hardware and single-class label vocabulary are not presented as new inventions. The contribution is an expanded industrial dataset accompanied by detailed quantitative characterization and a broader empirical benchmark.

The revised manuscript reports 25,120 native RGB line-scan images, comprising 12,654 defect-containing and 12,466 defect-free images, with 18,557 localized defect instances. It quantifies multiple-instance frequency, annotation geometry, boundary prevalence, and measured split characteristics. The experimental contribution includes three detector families, nine one-class anomaly-detection methods, a controlled four-resolution study, a boundary-subgroup evaluation, and a complete three-dataset zero-shot transfer matrix.

We have therefore clarified the contribution in terms of the documented dataset and the research questions supported by these analyses. We do not claim new hardware, imaging modalities, or a new defect taxonomy.

• The comparative table of public datasets only lists basic volume metrics but fails to conduct in-depth qualitative comparison with AITEX and RAW-FABRID (existing line-scan datasets). The authors do not quantitatively analyze how LUSITANOv2 outperforms prior line-scan benchmarks in defect diversity, imaging interference richness, or annotation quality, weakening the core argument of "largest and most comprehensive line-scan textile dataset".

We agree that image counts alone are insufficient to establish a dataset’s value. Table 1 now includes acquisition sensor, collection setting, colour, resolution, annotation format, fabric coverage, and released defect classes in addition to dataset size. We distinguish unverified properties from properties known to be absent.

The revised manuscript positions LUSITANOv2 as a complementary benchmark for active-production line-scan inspection. Its quantitative characterization includes defect-instance counts, relative defect size, overlapping annotations, and boundary prevalence. These measurements describe the released dataset but do not establish superiority over other datasets in defect diversity, interference richness, or annotation quality.

Accordingly, we have narrowed the comparative interpretation. We acknowledge that the remaining unverified entries limit the comparison and that the present evidence does not establish that LUSITANOv2 is the most comprehensive dataset across all relevant dimensions.

• Unreasonable input resolution and inconsistent experimental hyperparameter settings. The experiments use contradictory input resolutions (256×256 for general setup description vs. 512×128 actual input for line-scan images) without ablation studies to justify this compression choice. Line-scan fabric images are ultra-long horizontal strips; aggressive downsampling inevitably erases tiny subtle defects, yet the authors never analyze how resolution compression harms model performance, making benchmark results untrustworthy for real industrial deployment reference.

We agree that the preprocessing description needed to distinguish the network input canvas from the retained fabric content. Section 4.1 now explains that native 4096 × 1024 images are resized isotropically and padded to a square canvas. A 512 × 512 network input therefore contains a 512 × 128 fabric region. This preserves image aspect ratio while reducing spatial detail.

To quantify this reduction’s effect, we report four independently trained YOLOv12n models using canvas sizes of 256, 384, 512, and 640 pixels, corresponding to fabric regions of 256 × 64, 384 × 96, 512 × 128, and 640 × 160. The controlled runs use the same training schedule and seed, with early stopping disabled.

Increasing the canvas size from 256 to 640 improves mAP@0.50 from 0.449 to 0.598 and recall from 0.449 to 0.581. These results confirm that downsampling has a measurable cost. We distinguish this controlled study from the multi-architecture comparison and identify higher-resolution and native-patch processing as future directions. The results do not establish 512 pixels as an optimal deployment resolution.

• The core selling point of LUSITANOv2 is its real industrial domain shift challenge, yet the authors do not conduct cross-domain transfer experiments (training on lab datasets TILDA/ZJU-Leaper and testing on LUSITANOv2, or vice versa). There is no quantitative proof that models trained on lab data suffer severe performance drop on this dataset, which invalidates the central claim that the benchmark addresses industrial domain shift gaps.

We agree and have added a complete 3 × 3 zero-shot transfer experiment involving LUSITANOv2, TILDA, and ZJU-Leaper. One YOLOv12n model is trained on each dataset and evaluated on all three test sets. The protocol specifies a common architecture, input size, optimization recipe, and class-agnostic label mapping, without target-domain fine-tuning or adaptation.

Table 7 reports substantial transfer degradation. The TILDA-trained model achieves 0.805 mAP@0.5 on TILDA but 0.059 on LUSITANOv2. The ZJU-Leaper-trained model achieves 0.748 on ZJU-Leaper but 0.009 on LUSITANOv2. Transfer in the reverse direction is also weak: the LUSITANOv2-trained model obtains 0.035 on TILDA and 0.024 on ZJU-Leaper.

These results provide quantitative evidence of limited cross-dataset generalization. We explicitly acknowledge that the experiment does not isolate the effects of acquisition hardware, fabric composition, defect appearance, or other domain differences, and does not constitute an evaluation of a domain-adaptation method.

• The dataset does not store corresponding production parameters (fabric material, weaving type, production speed, lighting adjustment records, raw yarn specifications) captured during image acquisition. These metadata are critical for researching causal analysis of defect formation and domain adaptive algorithms, yet the authors discard all contextual factory data, limiting multi-modal and industrial causal research based on this dataset.

We agree that image-linked production metadata would support additional research, particularly process-conditioned analysis and investigation of defect formation. The current release provides images and bounding-box annotations but does not provide synchronized per-image records of material composition, weaving structure, production speed, lighting adjustments, or yarn specifications. The acquisition description in Section 3 documents the inspection setup but does not substitute for these contextual records.

This is a limitation of the present release: analyses requiring those variables cannot be conducted from the released images and annotations alone. At the same time, their absence from the release should not be interpreted as evidence that all contextual factory records were collected and subsequently discarded.

A useful future extension would link images to fabric or batch identifiers and synchronized production and acquisition records, subject to their availability and release permissions. Such an extension would broaden the benchmark’s scope beyond visual detection and transfer evaluation.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors created LUSITANOv2, a large real-world fabric defect dataset collected with industrial line-scan cameras from actual textile production lines. This dataset helps ease the domain shift issue seen in most lab-based textile datasets, and the paper also supplies full benchmark results for supervised object detection and unsupervised one-class anomaly detection. Even though the study has several meaningful innovations, there are still multiple points that need revision, as detailed below:
1.The abstract is too short and does not clearly state the core contributions of this work. I suggest expanding the abstract to give a clearer description of the paper’s novelties.
2.The literature review divides fabric defect detection algorithms into traditional methods and deep learning-based methods, and lists popular public datasets in a comparison table. However, the analysis of weaknesses of existing datasets is not thorough enough. The authors should add quantitative metrics like fabric variety, coverage of different defect types and real industrial interference levels, so readers can better understand the strengths of LUSITANOv2.
3.Only several classic detection and anomaly detection models are tested in this paper. Segmentation networks, lightweight models for factory deployment and domain adaptation approaches, which are widely adopted in textile inspection research nowadays, are not evaluated. The authors can either add a small set of supplementary experiments, or explain why these models are not tested in the discussion section.
4.The paper notes over 68% of defects are cut off at image edges, which is a distinctive difficulty of this dataset, yet it does not analyze how current models perform poorly on these incomplete defects. It is better to add visualized failure cases caused by edge-truncated defects, and compare the performance difference of each tested model on this hard sample group.
5.The discussion about dataset limitations and future research directions is too brief. The authors should expand this part, summarize other drawbacks of the current dataset (for example, all data are collected from one single factory, and the line-scan camera parameters remain unchanged), and propose more specific and practical follow-up research plans based on the benchmark data presented in this paper.

Author Response

We thank you for your detailed and constructive feedback. We've revised the manuscript to consider your criticisms, and we believe it's in a much better state now. We now provide a point-for-point response:

1.The abstract is too short and does not clearly state the core contributions of this work. I suggest expanding the abstract to give a clearer description of the paper’s novelties.

We have expanded the abstract to describe the acquisition setting, dataset composition, supported tasks, and principal experimental findings. It now reports the image and annotation counts, identifies the supervised and one-class benchmarks, and summarizes the resolution study, boundary-subgroup analysis, and cross-dataset transfer results.
These additions make the contribution and scope of the benchmark explicit.

2.The literature review divides fabric defect detection algorithms into traditional methods and deep learning-based methods, and lists popular public datasets in a comparison table. However, the analysis of weaknesses of existing datasets is not thorough enough. The authors should add quantitative metrics like fabric variety, coverage of different defect types and real industrial interference levels, so readers can better understand the strengths of LUSITANOv2.

We agree that readers need a comparison extending beyond image volume. We have expanded Table 1 to distinguish acquisition sensor, collection setting, resolution, annotation format, fabric coverage, and released defect classes. We have also expanded the quantitative characterization of LUSITANOv2, including defect multiplicity, relative annotation size, overlap, boundary prevalence, and measured split characteristics. We distinguish these measured properties from unverified fabric-taxonomy and interference information. Consequently, the revised discussion presents LUSITANOv2 as a complementary industrial benchmark without claiming quantified superiority in fabric variety or defect-category coverage. We acknowledge that unverified entries remain a limitation of the comparative analysis.

 

3.Only several classic detection and anomaly detection models are tested in this paper. Segmentation networks, lightweight models for factory deployment and domain adaptation approaches, which are widely adopted in textile inspection research nowadays, are not evaluated. The authors can either add a small set of supplementary experiments, or explain why these models are not tested in the discussion section.

We have expanded the explanation of the benchmark scope. The supervised comparison includes the lightweight one-stage model YOLOv12n alongside Faster R-CNN and RT-DETR-L, representing two-stage and transformer-based detection. YOLOv12n has 2.57 million parameters, compared with 43.26 million for Faster R-CNN and 31.99 million for RT-DETR-L.

We explain that supervised segmentation is not evaluated because the release lacks pixel-level ground truth. Although some evaluated anomaly-detection methods produce spatial anomaly maps, their reported evaluation is image-level AUROC and should not be interpreted as a supervised segmentation benchmark.

We also add zero-shot cross-dataset experiments to establish a baseline for future adaptation research, while explicitly acknowledging that no domain-adaptation algorithm is evaluated. Section 6 identifies matched latency, throughput, and memory measurements as necessary further work; parameter count alone does not establish deployment performance.

4.The paper notes over 68% of defects are cut off at image edges, which is a distinctive difficulty of this dataset, yet it does not analyze how current models perform poorly on these incomplete defects. It is better to add visualized failure cases caused by edge-truncated defects, and compare the performance difference of each tested model on this hard sample group.

We agree that boundary-related behaviour required direct evaluation. We first clarify the annotation-based definitions: exact boundary contact occurs in 65.08% of valid boxes, while 71.11% lie on or within one native pixel of a boundary. These definitions replace an ambiguous single prevalence figure. Boundary contact is an annotation-based proxy and does not independently establish physical defect truncation.

We compare exact-border and interior ground-truth recall for all three detectors on the same compatible test cohort, using confidence ≥ 0.50 and one-to-one matching at IoU ≥ 0.50. Border/interior recall is 0.437/0.098 for YOLOv12n, 0.793/0.567 for Faster R-CNN, and 0.747/0.432 for RT-DETR-L.

Contrary to the anticipated direction, boundary-touching boxes do not have lower recall at this operating point. We report this finding without interpreting boundary contact as causally beneficial, because the groups differ in geometry, size, location, and context. Figure 9 provides examples of successful detections, missed defects, and localization errors.

 

5.The discussion about dataset limitations and future research directions is too brief. The authors should expand this part, summarize other drawbacks of the current dataset (for example, all data are collected from one single factory, and the line-scan camera parameters remain unchanged), and propose more specific and practical follow-up research plans based on the benchmark data presented in this paper.

We have expanded Section 6 to identify the limitations of single-site acquisition, one inspection station, one camera placement, and a largely fixed illumination configuration. We also discuss the absence of fine-grained classes, severity labels, and pixel-level masks; the loss of detail under downsampling; confounding in boundary-subgroup comparisons; and the lack of matched deployment-efficiency measurements and evaluated adaptation methods.

The proposed research directions now follow directly from these findings: multi-site acquisition, richer annotations, native-resolution patches and rectangular processing, sequential or overlapping views for boundary context, domain adaptation and generalization, and accuracy–throughput evaluation on matched hardware.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have provided thorough responses to all my review comments. All raised issues have been well resolved. I do not have any additional questions and recommend the publication of this manuscript.