Multimodal Benchmark for Safety Assessment in Industrial Inspection Scenarios

Donghua Zhou; Jihai Min; Jun Cen; Pengyu Han; Shuang Liu; Songqiao Hu; Xiao He; Zeyi Liu; Zhaoheng Zhang; Zihan Meng

REVIEW 3 major objections 5 minor 44 references

Multimodal Benchmark for Safety Assessment in Industrial Inspection Scenarios

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read InspecSafe-V1: the first multimodal benchmark for industrial safety assessment built from real inspection-robot operations.

desk verdict A genuinely useful real-robot industrial inspection dataset with a solid annotation pipeline, but the benchmark evaluation is under-powered and the safety-label reliability is not yet demonstrated. read the letter →

arxiv 2601.21173 v2 pith:INBGGD7Z submitted 2026-01-29 cs.RO cs.CV

Zeyi Liu , Shuang Liu , Jihai Min , Zhaoheng Zhang , Jun Cen , Pengyu Han , Songqiao Hu , Zihan Meng

show 2 more authors

Xiao He Donghua Zhou

This is my paper · ORCID

classification cs.ROcs.CV

keywords industrialinspectionsafetyassessmentmultimodalbenchmarkvision-languagemodelsinstancesegmentationrobotsdatasetanomalydetection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces InspecSafe-V1, a dataset and benchmark for safety assessment in industrial inspection, collected from the routine operations of 41 deployed wheeled and rail-mounted inspection robots across five real industrial environments—tunnels, power facilities, sintering equipment, oil and gas petrochemical plants, and coal conveyor trestles. It comprises 5,013 inspection instances with pixel-level instance segmentation for 234 object categories, scene descriptions, four-level safety labels, and seven synchronized sensing modalities (RGB, thermal, audio, depth, radar point clouds, gas, temperature, humidity). The authors claim this is the first multimodal benchmark of its kind built entirely from real-world robot inspection data rather than simulations or lab setups, and they argue it closes the distribution gap that limits industrial AI deployment. Initial evaluations of general-purpose vision-language models on the benchmark show that reasoning-enhanced variants outperform instruction-only ones and that model size alone does not predict safety-assessment accuracy.

What carries the argument

The organizing unit is the 'inspection instance,' a short synchronized recording at a predefined inspection point, which ties all modalities together through inspection-point identifiers and timestamps. Annotations are generated by a pipeline that samples one frame every five frames, removes near-duplicates via deep-feature similarity, and then applies polygon-based instance segmentation plus structured language annotation of scene description and safety level. The safety level is assigned using a scenario-specific rule table (Table 3) that maps concrete hazards to Levels I–IV. For evaluation, the paper defines two metrics: safety-level prediction accuracy and semantic similarity between mod

What would settle it

Re-annotate a random subset of inspection instances (e.g., 200) with several independent annotators following Table 3 and measure inter-annotator agreement; if agreement is low (Cohen's kappa below 0.6) or if the re-annotations disagree with the published levels on more than a few percent of cases, the benchmark's ground truth and all VLM conclusions would be undermined.

Watch

Extended reading notes

Core claim

The central claim is that a large-scale, safety-annotated multimodal benchmark can be assembled from real inspection-robot operations and that this benchmark provides a meaningful test bed for vision-language models on industrial safety reasoning. Concretely, the paper asserts that InspecSafe-V1 is the first multimodal benchmark dataset for industrial inspection safety assessment collected from real-world environments, with each inspection instance carrying pixel-level segmentation, a semantic scene description, and a Level I–IV safety label, alongside temporally synchronized thermal, acoustic, point-cloud, gas, temperature, and humidity data. The dataset is released publicly, and the initia

Load-bearing premise

The benchmark's ground truth—safety level labels and scene descriptions—is taken as correct for scoring models, but it was checked on only a 10% sample of semantic annotations with a >95% internal accuracy threshold and no reported inter-annotator agreement, so label noise could shift the model rankings.

Editorial extensions

If this is right

The dataset gives industrial-AI researchers a standardized, reproducible benchmark for comparing vision-language models on safety assessment, currently missing from the field.
Because every instance pairs RGB with thermal, audio, point clouds, gas, temperature, and humidity data, it creates a testbed for cross-modal fusion and multimodal anomaly detection in real industrial conditions.
The inclusion of natural disturbances—illumination swings, reflections, occlusion, dust, and sensor noise—means models trained or evaluated on it are closer to deployment reality than on simulated or lab data.
The reported error patterns (scene misclassification and missed small-object violations) and false-positive differences give concrete guidance for choosing models, designing prompts, and setting alarm thresholds in practice.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the dataset is widely adopted, the four-level safety label could become the standard outcome variable for industrial foundation models, but the absence of inter-annotator agreement statistics means a public reliability study is a natural next step.
The synchronized non-visual modalities invite a testable extension the paper leaves implicit: using thermal trends or gas readings to predict Level I–II hazards earlier than visual cues alone would allow, which could be benchmarked as a separate task.
The strong performance of reasoning-augmented vision-language variants suggests a concrete hypothesis for future work: augmenting smaller models with explicit intermediate reasoning steps may close most of the gap to larger models, which the paper's data hint at but do not statistically prove.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know one thing up front: this paper is mostly a dataset release, and the dataset itself looks real and useful. InspecSafe-V1 gives the community 5,013 inspection instances from 41 real robots across five industrial sites, with RGB, thermal, audio, depth, radar, gas, and temperature/humidity, plus pixel-level instance masks, scene descriptions, and safety-level labels. That combination genuinely does not exist in the datasets they compare against (MVTec, Real-IAD, KITTI, nuScenes, ScanNet, COCO). The data collection protocol is described in enough detail to be reproducible, the file formats are standard, the data is publicly released, and the annotation pipeline (frame sampling, de-duplication, two-round QA with >95% thresholds) is clearly specified. That is real contribution, and I believe the 'first real-robot multimodal industrial inspection safety benchmark' claim is plausible even if they assert it rather than prove it exhaustively. Credit where due: the organization is careful, the limitation section honestly notes the discrete-label weakness, and the paper does not oversell the VLM results.

The soft spots are real but not fatal. The biggest one is the benchmark evaluation section: it reports accuracy and semantic similarity for a set of VLMs with zero baselines, no error bars, and no statistical testing. Figure 6 is essentially a table of point estimates. That is fine for a dataset paper — the dataset is the contribution, not the model ranking — but the text draws conclusions about model families (reasoning vs. instruction-tuned, false-positive rates) that the evaluation is not designed to support. The second soft spot is the safety-level ground truth. The QA protocol samples only 10% of semantic annotations and reports no inter-annotator agreement. Table 3's criteria include subjective calls (water pooling, using mobile phone, foreign objects), so label noise is a legitimate concern. The paper does not report the actual pass rates from the QA rounds, which would have helped. I do not think this invalidates the dataset, but it does mean the benchmark's central ground-truth reliability is unverified, and any model ranking built on it should be treated as provisional. A minor point: the 'first' claim is asserted rather than systematically evidenced, but I checked the comparison table and it holds up against the listed alternatives; I would not push on this much.

Who is this for? Researchers working on industrial anomaly detection, VLM safety reasoning, or multimodal fusion who want real-world, robot-centric data. It deserves a serious referee — the dataset is valuable enough that the evaluation weaknesses should be fixed through revision, not desk rejection.

My recommendation: send it to peer review. Ask the authors to add inter-annotator agreement, report QA pass rates, and present the VLM evaluation with error bars and a simple baseline (e.g., zero-shot CLIP or a random/constant classifier). Those are tractable revisions, and the underlying resource is worth the effort.

Referee Report

3 major / 5 minor

Summary. The paper introduces InspecSafe-V1, a multimodal benchmark dataset for safety assessment in industrial inspection scenarios. Data were collected from 41 wheeled and rail-mounted inspection robots at 2,239 inspection points across five industrial domains, yielding 5,013 inspection instances. Each instance includes pixel-level instance segmentation for visible-spectrum images, a semantic scene description, and a four-level safety label, alongside synchronized modalities such as thermal video, audio, depth and radar point clouds, gas, temperature, and humidity. The paper describes the data acquisition, annotation pipeline, quality-control protocol, and a benchmark evaluation of 15 vision-language models (VLMs) on safety-level prediction and semantic similarity of scene descriptions.

Significance. If validated, InspecSafe-V1 would fill a genuine gap: a real-world, multimodal, safety-annotated dataset for industrial inspection, enabling standardized evaluation of VLMs in safety-critical environments. The strengths are the real deployment provenance, multi-modal sensor suite, fine-grained pixel-level annotations, public release, and a structured two-round QA pipeline. However, the central claim as a benchmark rests on the reliability of the safety-level labels and the integrity of the evaluation protocol, both of which need stronger evidence. The dataset itself is a potentially valuable community resource, and the paper's contribution is significant if the identified validation gaps are addressed.

major comments (3)

[Technical Validation / Data Quality Assessment; Table 3] The safety-level labels are the core ground truth for the benchmark, but their reliability is not established. The QA protocol reports a two-round verification on 10% of semantic annotations with an internal >95% accuracy threshold, yet it does not report the actual pass rate, the number of samples reviewed, reviewer qualifications, or any inter-annotator agreement statistic (e.g., Cohen's kappa). Given the scenario-dependent and partly subjective criteria in Table 3 ('water pooling', 'foreign objects', 'using mobile phone'), different annotators may systematically assign different levels. Without label reliability estimates, the VLM accuracy comparisons in Fig. 6 and model rankings are unverified. Please add inter-annotator agreement, per-level annotation confusion, and details on how disagreements were resolved.
[Benchmark Evaluation, split construction (p. 8–9)] The train/test split is described as 'uniform intra-point sampling', which is ambiguous. If training and testing frames can originate from the same inspection-point video or the same inspection site, near-duplicate frames may leak information and inflate measured accuracy. The split must be performed at the inspection-point or video level, or at least the authors should demonstrate that no site contributes to both splits. This is a fundamental requirement for a trustworthy benchmark and directly affects the reported numbers.
[Benchmark Evaluation, Eq. (1), Eq. (2), Fig. 6] The benchmark reports single-point accuracy and semantic similarity for 15 VLMs without error bars, confidence intervals, or statistical significance tests. With N=1,250 and an imbalanced class distribution (999 normal vs 251 abnormal), the standard error of the accuracy estimate is roughly 1.4 percentage points, so differences of a few points (including the 'approximately eight percentage points' claim) may not be reliable. No baselines (e.g., majority-class, random, or a simple visual-only model) are reported. Please add bootstrap confidence intervals or pairwise significance tests (e.g., McNemar) to support the qualitative conclusions about model rankings and false-positive behavior.

minor comments (5)

[Figure 1] Typo: 'Inceptions Points' should be 'Inspection Points'. Also the affiliation at the title page has 'Beiging' (should be 'Beijing').
[Figure 5] The caption says 'Illustration of the key sensor module configurations...' but the figure shows data distributions (normal/abnormal ratio, object frequency, robot model distribution). The caption should match the content.
[Appendix S1] The prompt uses 'Level four' while the main text and Table 3 use 'Level IV'; standardize the naming for the no-abnormality class.
[Data Records / File Structure] The 'Other modalities' directory includes RGB video, but RGB is also part of the 'Annotations' directory; clarify the relationship between the two (e.g., raw video vs. extracted keyframes).
[Reference [45]] The arXiv number appears to have an extra digit: 'arXiv:2402.032164' should be 'arXiv:2402.03216'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: InspecSafe-V1's construction and evaluation are self-contained; author self-citations are background only.

full rationale

The paper does not derive any mathematical result from fitted parameters. Its central claim is the release of a multimodal dataset collected from real inspection robots; the ground-truth safety levels are human annotations produced according to the criteria in Table 3, not outputs of the VLMs being evaluated. The evaluation (Eq. 1 and Eq. 2) compares external VLM predictions to these fixed human labels, so accuracy/semantic similarity are not forced by construction. The QA protocol (5%/10% sampling with >95% accuracy thresholds) is an annotation quality-control step, not a fitted parameter renamed as a prediction; it does not define the benchmark's target labels in terms of the models' outputs. The authors' prior work (refs 2, 4, 6, 8, 11-13, 18) is cited only for motivation and background and is not invoked to validate the dataset or to forbid alternative benchmarks. No self-citation is load-bearing, no ansatz is smuggled via citation, and no known result is merely renamed. The paper's own Limitations section notes the discreteness of safety labels and limited scale, which are acknowledged scope limits rather than evidence of circularity. Therefore the benchmark construction and evaluation are self-contained.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No new theoretical entities or fitted constants are introduced; the paper is a dataset release and an empirical evaluation. The listed axioms are the unproved background assumptions the benchmark's validity rests on.

assumptions (4)

domain assumption Sampling one frame per five and the deep-learning deduplication step yield keyframes that faithfully represent each inspection point's content.
Used in Pipeline (Fig. 4); if sampled frames miss or over-represent objects, the 5,013 instances are not representative of the sites.
domain assumption The scenario-specific safety-level rubric (Table 3 / Appendix S1) is a valid operationalization of industrial safety, and annotators can apply it consistently.
All safety-level labels derive from this rubric; inconsistent application would corrupt the benchmark's ground truth.
domain assumption The quality-control protocol (5% sample for visual annotations, 10% for semantic annotations, >95% accuracy threshold) guarantees that annotation errors do not systematically bias the benchmark.
No inter-annotator agreement or error analysis is reported, so the stated QA sample sizes and threshold are assumed sufficient.
domain assumption BGE-M3 cosine similarity is a valid measure of semantic consistency between model-generated and human scene descriptions.
Used in Eq. (2) to compute SemSim; the validity of this metric as a quality measure is not independently justified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multimodal Benchmark for Safety Assessment in Industrial Inspection Scenarios." pith.science (2026). https://pith.science/paper/INBGGD7Z

@misc{pith2026260121173,
  author       = {Pith},
  title        = {Pith review of: Multimodal Benchmark for Safety Assessment in Industrial Inspection Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/INBGGD7Z}},
  note         = {Machine review of arXiv:2601.21173}
}

read the original abstract

With the rapid development of industrial intelligence and unmanned inspection, reliable perception and safety assessment for AI systems in complex and dynamic industrial sites has become a key bottleneck for deploying predictive maintenance and autonomous inspection. Most public datasets remain limited by simulated data sources, single-modality sensing, or the absence of fine-grained object-level annotations, which prevents robust scene understanding and multimodal safety reasoning for industrial foundation models. To address these limitations, InspecSafe-V1 is released as the first multimodal benchmark dataset for industrial inspection safety assessment that is collected from routine operations of real inspection robots in real-world environments. InspecSafe-V1 covers five representative industrial scenarios, including tunnels, power facilities, sintering equipment, oil and gas petrochemical plants, and coal conveyor trestles. The dataset is constructed from 41 wheeled and rail-mounted inspection robots operating at 2,239 valid inspection sites, yielding 5,013 inspection instances. For each instance, pixel-level segmentation annotations are provided for key objects in visible-spectrum images. In addition, a semantic scene description and a corresponding safety level label are provided according to practical inspection tasks. Seven synchronized sensing modalities are further included, including infrared video, audio, depth point clouds, radar point clouds, gas measurements, temperature, and humidity, to support multimodal anomaly recognition, cross-modal fusion, and comprehensive safety assessment in industrial environments.

Figures

Figures reproduced from arXiv: 2601.21173 by the authors.

**Figure 1.** Overview of InspecSafe-V1 Benchmark Dataset. valuable foundation for the development and evaluation of large multimodal models in industrial inspection. Methods Setup InspecSafe-V1 is constructed based on real-world industrial inspection tasks. Data collection relies on two types of on-site deployed robotic platforms: wheeled inspection robots and rail-mounted inspection robots. The two platforms exhibit complementa… view at source ↗

**Figure 2.** Illustration of the key sensor module configurations for wheeled and rail-mounted inspection platforms. scene are documented within the dataset. For the specific hardware configuration, the RGB camera is equipped with a 1/2.8-inch CMOS sensor, offering a field of view (FoV) ranging from 55.8° to 2.3° horizontally, 31.9° to 1.3° vertically, and 63.7° diagonally. The frame rate is 25 FPS. The thermal infrared camera e… view at source ↗

**Figure 3.** Illustration of the data acquisition process and sampling trajectories. At each inspection point, RGB video is stored in MP4 format with a resolution of either 2560×1440 or 1920×1080 pixels, and a recording duration of 10 to 15 seconds per clip. Thermal video is also saved in MP4 format, with a resolution of either 1280×1024 or 640×480 pixels, and the same recording duration of 10 to 15 seconds. Environmental audio … view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Overview of the InspecSafe-V1 dataset construction pipeline. During the construction and execution of the annotation protocol, labeling scope is defined based on objects of interest in industrial inspection tasks. A structured and extensible label taxonomy is establish…

**Figure 5.** Figure 5: Illustration of the key sensor module configurations for wheeled and rail-mounted inspection platforms. Prediction correctness is determined by comparing yˆi against the ground truth label yi . Referred to the literature44, the utilized prompt for benchmark evaluation …

**Figure 6.** Figure 6: Benchmark results of different VLMs on InspecSafe-V1. Reasoning-oriented models consistently achieve better performance compared to their instruction-only counterparts. Within the Qwen3-VL family, the reasoning-enhanced variant achieves approximately eight percentage p…

**Figure 7.** Figure 7: Qualitative comparison of safety assessment results across different VLMs. 15/15 [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 4 linked inside Pith

[1]

D., Rodd, M

Thomas, A. D., Rodd, M. G., Holt, J. D. & Neill, C. Real-time industrial visual inspection: A review.Real-Time Imaging 1, 139–158 (1995)

1995
[2]

& Liu, Z

He, X. & Liu, Z. Dynamic model interpretation-guided online active learning scheme for real-time safety assessment. IEEE Transactions on Cybern.54, 2734–2745, DOI: 10.1109/TCYB.2023.3339242 (2024)

arXiv 2023
[3]

& Shpitalni, M

Weimer, D., Scholz-Reiter, B. & Shpitalni, M. Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection.CIRP annals65, 417–420 (2016)

2016
[4]

& Zhou, D

He, X., Liu, Z., Hu, S., Liu, C. & Zhou, D. Real-time safety assessment techniques of dynamic systems.Acta Autom. Sinica51, 249–270 (2025)

2025
[5]

& Ning, Z

Gao, Y ., Lin, J., Xie, J. & Ning, Z. A real-time defect detection method for digital signal processing of industrial inspection applications.IEEE Transactions on Ind. Informatics17, 3450–3459 (2020). 3https://github.com/CVHub520/X-AnyLabeling 11/15

2020
[6]

Liu, Z. & He, X. Online dynamic hybrid broad learning system for real-time safety assessment of dynamic systems.IEEE Transactions on Knowl. Data Eng.36, 8928–8938, DOI: 10.1109/TKDE.2024.3475028 (2024)

arXiv 2024
[7]

Data12, 418 (2025)

Khandakar, A.et al.Harnessing smartphone sensors for enhanced road safety: A comprehensive dataset and review.Sci. Data12, 418 (2025)

2025
[8]

Liu, Z., Hu, S. & He, X. Real-time safety assessment of dynamic systems in non-stationary environments: A review of methods and techniques. In2023 CAA Symposium on Fault Detection, Supervision and Safety for Technical Processes (SAFEPROCESS), 1–6 (IEEE, 2023)

2023

Show all 44 references

[9]

In2019 25th International Conference on Automation and Computing (ICAC), 1–6 (IEEE, 2019)

Yu, L.et al.Inspection robots in oil and gas industry: a review of current solutions and future trends. In2019 25th International Conference on Automation and Computing (ICAC), 1–6 (IEEE, 2019)

2019
[10]

& De Almeida, A

Tavakoli, M., Marques, L. & De Almeida, A. T. Development of an industrial pipeline inspection robot.Ind. Robot: An Int. J.37, 309–322 (2010)

2010
[11]

Liu, Z. & He, X. Dynamic submodular-based learning strategy in imbalanced drifting streams for real-time safety assessment in nonstationary environments.IEEE Transactions on Neural Networks Learn. Syst.35, 3038–3051 (2024)

2024
[12]

Liu, Z., Zhang, Y ., Ding, Z. & He, X. An online active broad learning approach for real-time safety assessment of dynamic systems in nonstationary environments.IEEE Transactions on Neural Networks Learn. Syst.34, 6714–6724, DOI: 10.1109/TNNLS.2022.3222265 (2023)

2022
[13]

Hu, S., Liu, Z., Li, M. & He, X. CADM+: Confusion-based learning framework with drift detection and adaptation for real-time safety assessment.IEEE Transactions on Neural Networks Learn. Syst.36, 5126–5139, DOI: 10.1109/TNNLS. 2024.3369315 (2025)

2024
[14]

& Jough, F

Soori, M., Dastres, R., Arezoo, B. & Jough, F. K. G. Intelligent robotic systems in industry 4.0: A review.J. Adv. Manuf. Sci. Technol.2024007–0 (2024)

2024
[15]

& Gan, Z

Mei, A., Zhu, G.-N., Zhang, H. & Gan, Z. Replanvlm: Replanning robotic tasks with visual language models.IEEE Robotics Autom. Lett.(2024)

2024
[16]

Zeng, F., Gan, W., Wang, Y ., Liu, N. & Yu, P. S. Large language models for robotics: A survey.arXiv preprint arXiv:2311.07226(2023)

2023
[17]

Data12, 117 (2025)

Wang, Y .et al.A multi-modal dental dataset for semi-supervised deep learning image segmentation.Sci. Data12, 117 (2025). 18.Hu, S.et al.Vlsa: Vision-language-action models with plug-and-play safety constraint layer (2025). 2512.11891

2025 arXiv
[19]

& Kivist¨o-Rahnasto, J

Lind, S., Nenonen, S. & Kivist¨o-Rahnasto, J. Safety risk assessment in industrial maintenance.J. Qual. Maintenance Eng. 14, 205–217 (2008). 20.Le Coze, J.-c. Outlines of a sensitising model for industrial safety assessment.Saf. science51, 187–201 (2013)

2008
[21]

Critical reviews microbiology40, 187–206 (2014)

Groenewald, M.et al.Yarrowia lipolytica: safety assessment of an oleaginous yeast with a great industrial potential. Critical reviews microbiology40, 187–206 (2014)

2014
[22]

& Nie, Y

Hao, M. & Nie, Y . Hazard identification, risk assessment and management of industrial system: Process safety in mining industry.Saf. science154, 105863 (2022)

2022
[23]

& Kwolek, B

Wozniak, P., Krzeszowski, T. & Kwolek, B. Multi-domain indoor dataset for visual place recognition and anomaly detection by mobile robots.Sci. Data12, 817 (2025)

2025
[24]

& Steger, C

Bergmann, P., Fauser, M., Sattlegger, D. & Steger, C. Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9592–9600 (2019)

2019
[25]

& Dabeer, O

Zou, Y ., Jeong, J., Pemula, L., Zhang, D. & Dabeer, O. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. InEuropean conference on computer vision, 392–408 (Springer, 2022)

2022
[26]

In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 22883–22892 (2024)

Wang, C.et al.Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 22883–22892 (2024)

2024
[27]

InProceedings of the Computer Vision and Pattern Recognition Conference, 15214–15223 (2025)

Zhu, W.et al.Real-iad d3: A real-world 2d/pseudo-3d/3d dataset for industrial anomaly detection. InProceedings of the Computer Vision and Pattern Recognition Conference, 15214–15223 (2025)

2025
[28]

& Steger, C

Bergmann, P., Jin, X., Sattlegger, D. & Steger, C. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization.arXiv preprint arXiv:2112.09045(2021). 12/15

2021 arXiv
[29]

Neural Inf

Liu, J.et al.Real3d-ad: A dataset of point cloud anomaly detection.Adv. Neural Inf. Process. Syst.36, 30402–30415 (2023)

2023
[30]

& Siami, M

Behzad, K., Zandi, R., Motamedi, E., Salehinejad, H. & Siami, M. Robomnist: A multimodal dataset for multi-robot activity recognition using wifi sensing, video, and audio.Sci. Data12, 326 (2025)

2025
[31]

& Urtasun, R

Geiger, A., Lenz, P., Stiller, C. & Urtasun, R. Vision meets robotics: The kitti dataset.The international journal robotics research32, 1231–1237 (2013)

2013
[32]

InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11621–11631 (2020)

Caesar, H.et al.nuscenes: A multimodal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11621–11631 (2020)

2020
[33]

InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2446–2454 (2020)

Sun, P.et al.Scalability in perception for autonomous driving: Waymo open dataset. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2446–2454 (2020)

2020
[34]

Choi, Y .et al.Kaist multi-spectral day/night data set for autonomous and assisted driving.IEEE Transactions on Intell. Transp. Syst.19, 934–948 (2018)

2018
[35]

InProceedings of the IEEE conference on computer vision and pattern recognition, 5828–5839 (2017)

Dai, A.et al.Scannet: Richly-annotated 3d reconstructions of indoor scenes. InProceedings of the IEEE conference on computer vision and pattern recognition, 5828–5839 (2017)

2017
[36]

Song, S., Lichtenberg, S. P. & Xiao, J. Sun rgb-d: A rgb-d scene understanding benchmark suite. InProceedings of the IEEE conference on computer vision and pattern recognition, 567–576 (2015)

2015
[37]

& Berg, A

Ammirato, P., Poirson, P., Park, E., Koˇseck´a, J. & Berg, A. C. A dataset for developing and benchmarking active vision. In 2017 IEEE international conference on robotics and automation (ICRA), 1378–1385 (IEEE, 2017)

2017
[38]

InEuropean conference on computer vision, 740–755 (Springer, 2014)

Lin, T.-Y .et al.Microsoft coco: Common objects in context. InEuropean conference on computer vision, 740–755 (Springer, 2014)

2014
[39]

journal computer vision128, 1956–1981 (2020)

Kuznetsova, A.et al.The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale.Int. journal computer vision128, 1956–1981 (2020)

1956
[40]

journal computer vision123, 32–73 (2017)

Krishna, R.et al.Visual genome: Connecting language and vision using crowdsourced dense image annotations.Int. journal computer vision123, 32–73 (2017)

2017
[41]

Shinde, G.et al.A survey on efficient vision-language models.Wiley Interdiscip. Rev. Data Min. Knowl. Discov.15, e70036 (2025). 42.Bordes, F.et al.An introduction to vision-language modeling.arXiv preprint arXiv:2405.17247(2024)

2025 arXiv
[43]

neural information processing systems36, 49250–49267 (2023)

Dai, W.et al.Instructblip: Towards general-purpose vision-language models with instruction tuning.Adv. neural information processing systems36, 49250–49267 (2023)

2023
[44]

Zhou, K., Yang, J., Loy, C. C. & Liu, Z. Learning to prompt for vision-language models.Int. J. Comput. Vis.130, 2337–2348 (2022)

2022
[45]

Chen, J.et al.Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self- knowledge distillation.arXiv preprint arXiv:2402.032164(2024). Acknowledgements This work was supported in part by National Natural Science Foundation of China u...

2024 arXiv
[46]

Describe key visual elements in the image (e.g., personnel behavior, environmental conditions, object distribution, hazard sources, etc.)
[47]

Unrecognizable

Assess the safety level of the scene into one of four categories: - Level one: High safety threats present that could lead to severe incidents (e.g., open flame, smoke, personnel collapse, oil leakage, etc.). 13/15 - Level two: Moderate safety threats present that may cause sa...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

D., Rodd, M

Thomas, A. D., Rodd, M. G., Holt, J. D. & Neill, C. Real-time industrial visual inspection: A review.Real-Time Imaging 1, 139–158 (1995)

1995

[2] [2]

& Liu, Z

He, X. & Liu, Z. Dynamic model interpretation-guided online active learning scheme for real-time safety assessment. IEEE Transactions on Cybern.54, 2734–2745, DOI: 10.1109/TCYB.2023.3339242 (2024)

arXiv 2023

[3] [3]

& Shpitalni, M

Weimer, D., Scholz-Reiter, B. & Shpitalni, M. Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection.CIRP annals65, 417–420 (2016)

2016

[4] [4]

& Zhou, D

He, X., Liu, Z., Hu, S., Liu, C. & Zhou, D. Real-time safety assessment techniques of dynamic systems.Acta Autom. Sinica51, 249–270 (2025)

2025

[5] [5]

& Ning, Z

Gao, Y ., Lin, J., Xie, J. & Ning, Z. A real-time defect detection method for digital signal processing of industrial inspection applications.IEEE Transactions on Ind. Informatics17, 3450–3459 (2020). 3https://github.com/CVHub520/X-AnyLabeling 11/15

2020

[6] [6]

Liu, Z. & He, X. Online dynamic hybrid broad learning system for real-time safety assessment of dynamic systems.IEEE Transactions on Knowl. Data Eng.36, 8928–8938, DOI: 10.1109/TKDE.2024.3475028 (2024)

arXiv 2024

[7] [7]

Data12, 418 (2025)

Khandakar, A.et al.Harnessing smartphone sensors for enhanced road safety: A comprehensive dataset and review.Sci. Data12, 418 (2025)

2025

[8] [8]

Liu, Z., Hu, S. & He, X. Real-time safety assessment of dynamic systems in non-stationary environments: A review of methods and techniques. In2023 CAA Symposium on Fault Detection, Supervision and Safety for Technical Processes (SAFEPROCESS), 1–6 (IEEE, 2023)

2023