| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "arrayShape": "cr:arrayShape", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "containedIn": "cr:containedIn", |
| "cr": "http://mlcommons.org/croissant/", |
| "data": { |
| "@id": "cr:data", |
| "@type": "@json" |
| }, |
| "dataBiases": "cr:dataBiases", |
| "dataCollection": "cr:dataCollection", |
| "dataType": { |
| "@id": "cr:dataType", |
| "@type": "@vocab" |
| }, |
| "dct": "http://purl.org/dc/terms/", |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileProperty": "cr:fileProperty", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "format": "cr:format", |
| "includes": "cr:includes", |
| "isArray": "cr:isArray", |
| "isLiveDataset": "cr:isLiveDataset", |
| "jsonPath": "cr:jsonPath", |
| "key": "cr:key", |
| "md5": "cr:md5", |
| "parentField": "cr:parentField", |
| "path": "cr:path", |
| "personalSensitiveInformation": "cr:personalSensitiveInformation", |
| "recordSet": "cr:recordSet", |
| "references": "cr:references", |
| "regex": "cr:regex", |
| "repeated": "cr:repeated", |
| "replace": "cr:replace", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "prov": "http://www.w3.org/ns/prov#" |
| }, |
| "@type": "sc:Dataset", |
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "repo", |
| "name": "repo", |
| "description": "The Hugging Face git repository.", |
| "contentUrl": "https://huggingface.co/datasets/nvidia/PhysicalAI-VANTAGE-Bench/tree/refs%2Fconvert%2Fparquet", |
| "encodingFormat": "git+https", |
| "sha256": "https://github.com/mlcommons/croissant/issues/80" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "parquet-files-for-config-default", |
| "containedIn": { |
| "@id": "repo" |
| }, |
| "encodingFormat": "application/x-parquet", |
| "includes": "default/*/*.parquet" |
| } |
| ], |
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "dataType": "cr:Split", |
| "key": { |
| "@id": "default_splits/split_name" |
| }, |
| "@id": "default_splits", |
| "name": "default_splits", |
| "description": "Splits for the default config.", |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "default_splits/split_name", |
| "dataType": "sc:Text" |
| } |
| ], |
| "data": [ |
| { |
| "default_splits/split_name": "train" |
| } |
| ] |
| }, |
| { |
| "@type": "cr:RecordSet", |
| "@id": "default", |
| "description": "nvidia/PhysicalAI-VANTAGE-Bench - 'default' subset", |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "default/split", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "parquet-files-for-config-default" |
| }, |
| "extract": { |
| "fileProperty": "fullpath" |
| }, |
| "transform": { |
| "regex": "default/(?:partial-)?(train)/.+parquet$" |
| } |
| }, |
| "references": { |
| "field": { |
| "@id": "default_splits/split_name" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "default/image", |
| "dataType": "sc:ImageObject", |
| "source": { |
| "fileSet": { |
| "@id": "parquet-files-for-config-default" |
| }, |
| "extract": { |
| "column": "image" |
| }, |
| "transform": { |
| "jsonPath": "bytes" |
| } |
| } |
| } |
| ] |
| } |
| ], |
| "conformsTo": "http://mlcommons.org/croissant/1.1", |
| "name": "PhysicalAI-VANTAGE-Bench", |
| "description": "\n\t\n\t\t\n\t\tVANTAGE-BENCH\n\t\n\nVideo ANalysis Tasks Across Generalized Environments\n\n\t\n\t\t\n\t\tDataset Description\n\t\n\nVANTAGE-BENCH is the first public benchmark purpose-built for evaluating visual understanding on video captured by fixed infrastructure cameras. It spans three real-world domains \u2014 warehouse, smart city / Intelligent Transportation Systems (ITS), and smart spaces \u2014 across six spatio-temporal video understanding tasks including video question answering (VQA), temporal grounding, dense\u2026 See the full description on the dataset page: https://huggingface.co/datasets/nvidia/PhysicalAI-VANTAGE-Bench.", |
| "alternateName": [ |
| "nvidia/PhysicalAI-VANTAGE-Bench" |
| ], |
| "creator": { |
| "@type": "Organization", |
| "name": "NVIDIA", |
| "url": "https://huggingface.co/nvidia" |
| }, |
| "keywords": [ |
| "other", |
| "1K - 10K", |
| "imagefolder", |
| "Image", |
| "Video", |
| "Datasets", |
| "Croissant", |
| "arxiv:2502.00392", |
| "\ud83c\uddfa\ud83c\uddf8 Region: US" |
| ], |
| "license": "https://huggingface.co/datasets/nvidia/PhysicalAI-VANTAGE-Bench/blob/main/LICENSE", |
| "url": "https://huggingface.co/datasets/nvidia/PhysicalAI-VANTAGE-Bench", |
| "rai:dataLimitations": "VANTAGE-Bench is an evaluation-only benchmark; ground-truth annotations are held server-side and the dataset is not intended for model training or fine-tuning. Coverage is restricted to fixed-camera infrastructure footage in three operational domains and predominantly daylight, fair-weather conditions; nighttime, infrared, heavy snow, and other adverse-weather scenes are out of scope. Audio is not evaluated. The benchmark is single-view: no multi-camera correlation tasks are included. Annotations and supporting prompts are English-only. Anonymization, while audited, is best-effort: small, occluded, or motion-blurred faces and license plates may not have been detected. Strong performance on this benchmark is not evidence of clinical, legal, evidentiary, or safety-critical deployment readiness, and the benchmark should not be used for identity recognition, real-time closed-loop control, or any task requiring multi-camera tracking.", |
| "rai:dataBiases": "Real-world footage is concentrated in the United States: warehouse and smart-spaces clips come from US-based contracted vendor facilities, and transportation clips come from a single municipal partner (City of Dubuque, Iowa). The benchmark therefore reflects US infrastructure, signage, and traffic conventions. A portion of VQA / Temporal / DVC clips are synthetic (NVIDIA DriveSim Omniverse), and the single-object-tracking subset is fully synthetic (AI City Challenge Track 1), introducing a sim-to-real distribution shift the benchmark explicitly measures. The 2D Referring Expressions subset is fully aerial drone imagery (RefDrone) and is therefore not representative of ground-mounted infrastructure cameras. The Event Verification subset intentionally includes both positive event examples and plausible negatives; the precise per-class distribution is held out as part of the evaluation. Daylight and clear-weather scenes dominate every track. The 2D Spatial Pointing task is procedurally generated from human-verified bounding boxes, so its distribution inherits the bounding-box source's spatial bias.", |
| "rai:personalSensitiveInformation": "Source footage was captured in semi-public spaces (warehouses, traffic intersections, smart-space environments) and incidentally contains people and vehicles. Faces and vehicle license plates have been obfuscated by an automated de-identification pipeline. The released benchmark contains no direct personally identifiable information (PII) \\u2014 no names, no audio transcripts, no demographic metadata, no health or financial data. Geographic identifiers (intersection or facility names) appear in some filenames. Any clips originally sourced from third-party YouTube re-uploads are not redistributed; users obtain those from the original source themselves. An ethics review was conducted by NVIDIA in addition to the technical anonymization step.", |
| "rai:dataUseCases": "VANTAGE-Bench is designed for (1) zero-shot evaluation of vision-language models on fixed-camera ('Observer AI') video understanding across spatial, temporal, and semantic reasoning; (2) measuring the 'Observer AI Gap' between model performance on consumer / cinematic video and on real fixed-infrastructure footage; (3) measuring the sim-to-real gap by comparing accuracy on DriveSim-generated and real-world matched scenes; and (4) studying stateless, single-pass VLM tracking via VANTAGE-SOT \\u2014 the first quantitative evaluation protocol for VLMs as direct trackers. Validated tasks and metrics: VQA (Top-1 accuracy), Event Verification (macro F1), Temporal Localization (mIoU + Precision@0.5), Dense Video Captioning (SODA_c), 2D Object Localization (COCO mAP / F1), 2D Referring Expressions (Acc@IoU), 2D Spatial Pointing (pointing accuracy), Single-Object Tracking (mean spatial IoU + success AUC). The benchmark is not validated for and should not be used for model training, multi-camera reasoning, audio-dependent tasks, real-time control, identity recognition, or any clinical, legal, or safety-critical decision support.\"", |
| "rai:dataSocialImpact": "VANTAGE-Bench enables, for the first time, language-grounded evaluation of fixed-infrastructure video understanding across warehouse safety, transportation ITS, and smart spaces in one unified benchmark. By revealing the 'Observer AI Gap' (the measured drop in performance when frontier VLMs move from consumer video to fixed-infrastructure footage), it discourages premature claims of operational readiness for VLM-based safety systems. Foreseeable misuse risks include (a) using the released clips and visible annotations to train or fine-tune model weights, which would contaminate the held-out evaluation and invalidate published benchmark scores; (b) repurposing models tuned against this benchmark for surveillance against unredacted live feeds; (c) treating benchmark scores as evidence of deployment readiness for automated insurance, policing, or HR / disciplinary decisions; and (d) over-trusting strong sim-only performance on the synthetic tracking subset. Mitigations: the benchmark is evaluation-only with server-side held-out ground truth (preventing use as a training set); the dataset repository is gated, so users must accept the terms of use before downloading; the license explicitly prohibits training, fine-tuning, or any use other than evaluation; automated face and license-plate anonymization; and explicit limitations and use-case boundaries published with the benchmark.", |
| "rai:hasSyntheticData": true, |
| "prov:wasDerivedFrom": [ |
| { |
| "@id": "https://huggingface.co/datasets/nvidia/PhysicalAI-VANTAGE-Bench", |
| "prov:label": "nvidia/PhysicalAI-VANTAGE-Bench", |
| "sc:license": "NVIDIA EVALUATION DATA LICENSE" |
| }, |
| { |
| "@id": "https://arxiv.org/abs/2502.00392" |
| } |
| ], |
| "prov:wasGeneratedBy": [ |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q4929239" |
| }, |
| "prov:label": "Transportation data acquisition (City of Dubuque)", |
| "sc:description": "Fixed-camera highway / ITS footage acquired under a data-use agreement with the City of Dubuque, Iowa. The agreement permits redistribution as part of VANTAGE-Bench. Coverage spans the city's deployed traffic-camera network.", |
| "prov:wasAttributedTo": [ |
| { |
| "@type": "prov:Agent", |
| "@id": "city_of_dubuque,_iowa", |
| "prov:label": "City of Dubuque, Iowa", |
| "sc:description": "University and City collected Traffic Camera Data" |
| } |
| ] |
| }, |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q4929239" |
| }, |
| "prov:label": "Warehouse and smart-spaces footage collection", |
| "sc:description": "GoPro fixed-camera recordings collected at multiple US-based warehouse and smart-spaces facilities by contracted vendors. Subjects were informed and consented to recording; vendor agreements authorize redistribution under VANTAGE-Bench." |
| }, |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q4929239" |
| }, |
| "prov:label": "Synthetic Warehouse Data", |
| "sc:description": "A portion of VQA / Temporal / DVC clips were rendered on the NVIDIA DriveSim Omniverse simulator (collision and multi-camera scenarios). Synthetic data is included intentionally to allow direct sim-to-real comparison." |
| }, |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q1172378" |
| }, |
| "prov:label": "Face and license-plate anonymization", |
| "sc:description": "All real-world footage and frames were processed by an automated face- and license-plate-detection pipeline that applied mosaic blur to detected regions. Residual risk: small or occluded faces / plates may remain visible.", |
| "prov:wasAttributedTo": [ |
| { |
| "@type": "prov:SoftwareAgent", |
| "@id": "nvidia_face_and_plate_anonymization_pipeline", |
| "prov:label": "NVIDIA face-and-plate anonymization pipeline", |
| "sc:description": "NVIDIA ip to detect face-and-plate and anonymize PII" |
| } |
| ] |
| }, |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q109719325" |
| }, |
| "prov:label": "Acquiring Transformed Labeled Dataset", |
| "sc:description": "Curation including selection of clips around safety-critical incidents (collisions, tailgating, near-misses, zone breaches) and trimming of long recordings into evaluable clips was performed by the research authors using domain-specific rulebooks. Annotations for VQA, Event Verification, Temporal Localization, Dense Video Captioning, and 2D Object Localization were authored by trained NVIDIA Data Factory professionals (not crowdsourced). Annotators worked from domain-specific rulebooks defining safety-critical incident boundaries and spatial coordinate conventions. A secondary QA expert verified every bounding box, temporal segment, and caption before inclusion. Ten percent of human-annotated samples were independently cross-validated by a separate QA expert to establish construct validity, label consistency, and label reliability. The 2D Spatial Pointing (VANTAGE-2DPoint) task is generated by a pseudo-labeling pipeline that converts human-verified bounding boxes into spatial-reasoning question-answer pairs via templated heuristics over relative position, distance, and overlap.", |
| "prov:wasAttributedTo": [ |
| { |
| "@type": "prov:Agent", |
| "@id": "research_team", |
| "prov:label": "Research Team", |
| "sc:description": "Team clipped data and pseudolabeled data to get it into format for evaluation" |
| }, |
| { |
| "@type": "prov:Agent", |
| "@id": "nvidia_data_factory", |
| "prov:label": "NVIDIA Data Factory", |
| "sc:description": "Trained Professional Annotators executed guideline for labels" |
| }, |
| { |
| "@type": "prov:SoftwareAgent", |
| "@id": "vantage_2dpoint_pseudo_labeling_pipeline", |
| "prov:label": "VANTAGE-2DPoint pseudo-labeling pipeline", |
| "sc:description": "pseudo-labeling pipeline that converts human-verified bounding boxes into spatial-reasoning question-answer pairs" |
| } |
| ] |
| } |
| ], |
| "citeAs": "Bhat, Z. P., Nayyar, N., Jain, A., Chan, L. F., Suchanek, J., Wang, Y., Praveen, V., Kornuta, T., & Murali, V. N. (2026). VANTAGE-Bench: Evaluating the Infrastructure AI Gap in Vision-Language Models. In Advances in Neural Information Processing Systems (NeurIPS) Evaluations & Datasets Track.", |
| "citation": "Bhat, Z. P., Nayyar, N., Jain, A., Chan, L. F., Suchanek, J., Wang, Y., Praveen, V., Kornuta, T., & Murali, V. N. (2026). VANTAGE-Bench: Evaluating the Infrastructure AI Gap in Vision-Language Models. In Advances in Neural Information Processing Systems (NeurIPS) Evaluations & Datasets Track.", |
| "version": "1.0.0", |
| "datePublished": "2026-04-24" |
| } |
