Hugging Face's logo

Datasets:
nvidia
/
PhysicalAI-VANTAGE-Bench

Modalities:
Image
Tabular
Text
Size:
1K - 10K
ArXiv:
Tags:
video
video-understanding
benchmark
evaluation
infrastructure-cameras
warehouse
License:
Dataset card Data Studio Files
xet
 Community
25
PhysicalAI-VANTAGE-Bench / README.md
lapchann's picture
lapchann
polish readme
7dae046
preview code
|
Raw
History Blame
13.4 kB
metadata 
license: other
license_name: nvidia-evaluation-data-license
license_link: LICENSE
language:
  - en
pretty_name: VANTAGE-Bench
size_categories:
  - 10K<n<100K
task_categories:
  - visual-question-answering
  - video-text-to-text
  - image-text-to-text
  - object-detection
  - multiple-choice
task_ids:
  - visual-question-answering
  - image-captioning
  - multiple-choice-qa
tags:
  - video
  - image
  - text
  - multimodal
  - video-understanding
  - image-understanding
  - benchmark
  - evaluation
  - infrastructure-cameras
  - warehouse
  - smart-city
  - intelligent-transportation-systems
  - smart-spaces
configs:
  - config_name: vqa
    data_files:
      - split: test
        path: data/vqa/data_jsons/annotations/*.json
  - config_name: temporal_localization
    data_files:
      - split: test
        path: data/temporal_localization/data_jsons/annotations/*.json
  - config_name: event_verification
    data_files:
      - split: test
        path: data/event_verification/data_jsons/annotations/*.json
  - config_name: referring
    data_files:
      - split: test
        path: data/referring/refdrone_test_llava.json
  - config_name: pointing
    data_files:
      - split: test
        path: data/pointing/VANTAGE_2DPointing.jsonl
  - config_name: tracking
    data_files:
      - split: test
        path: data/tracking/sot_benchmark.jsonl
  - config_name: 2dbbox
    data_files:
      - split: test
        path: data/2dbbox/metadata.jsonl
  - config_name: dense_captioning
    data_files:
      - split: test
        path: data/dense_captioning/metadata.jsonl

VANTAGE-Bench

Video ANalysis Tasks Across Generalized Environments

3 domains · 8 tasks · 35,027 annotations · 3,346 media samples · 42 GB

VANTAGE-Bench task overview across Semantic, Temporal, Spatial, and Spatio-Temporal understanding categories

Dataset Description

VANTAGE-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 — warehouse, smart city / Intelligent Transportation Systems (ITS), and smart spaces — across 8 tasks spanning semantic, temporal, spatial, and spatio-temporal evaluation, including video question answering (VQA), temporal localization, dense video captioning (DVC), event verification, spatial pointing, referring expressions, and spatio-temporal tracking. Unlike ordinary web video, this footage comes from fixed, infrastructure-mounted viewpoints — persistent scenes under long-duration monitoring that demand reasoning over stationary warehouse, ITS, and smart-space environments.

Evaluation-only / test split. Ground-truth answers are withheld and all scoring is performed server-side — this dataset repository does not provide local benchmark scoring.

Directory Structure 

PhysicalAI-VANTAGE-Bench/
├── data/
│   ├── 2dbbox/                # 2D object localization
│   ├── dense_captioning/      # Dense video captioning
│   ├── event_verification/    # Event verification
│   ├── pointing/              # 2D spatial pointing
│   ├── referring/             # 2D referring expressions
│   ├── temporal_localization/ # Temporal localization
│   ├── tracking/              # Single object tracking
│   └── vqa/                   # Video question answering
│
├── scripts/
│   ├── run_lmudata.py         # Prepare benchmark datasets
│   └── RUN_LMUData.md         # Setup and usage guide
│
└── README.md                  # Dataset documentation

Get Started

This repository contains the official VANTAGE-Bench dataset and data schemas. For benchmark documentation, submissions, and leaderboard results, use the resources below:

Quick Start

This repository ships the test-split media and question-side annotations; ground-truth answers are withheld for server-side scoring. VANTAGE-Bench's evaluation toolkit expects benchmark datasets to be organized using a standard directory structure called LMUData. To build an inference-ready LMUData layout across every task:

python scripts/run_lmudata.py --all --lmu-root ~/LMUData

Run from a clone of this dataset repo, the script auto-uses the local data/ folder; otherwise it downloads the public dataset from Hugging Face. Two tasks fetch external media during this step — 2D Referring Expressions (RefDrone / VisDrone images) and Single Object Tracking (PhysicalAI-SmartSpaces videos; needs ffmpeg and an HF token). See scripts/RUN_LMUData.md for full setup, disk requirements, per-task notes, and troubleshooting.

What the setup produces

run_lmudata.py automates the inference-prep step end to end. It sources the public dataset (an auto-detected local data/ clone, an explicit --local-source, or a Hugging Face snapshot), builds each task's index file (*.tsv / annotations.json), and places the media by symlink (default) or --copy. It writes no ground-truth fields — withheld answers are left empty — and is idempotent, so re-runs only fill in what is missing.

Most tasks need nothing beyond the command above. Two have extra prerequisites, which the script handles automatically when they are met:

  • Single Object Tracking — downloads source videos from nvidia/PhysicalAI-SmartSpaces and extracts frames with ffmpeg; needs an HF token with read access to that (gated) dataset.
  • 2D Referring Expressions (grounding) — downloads the RefDrone / VisDrone images over the network.

Under --all, a task that cannot meet its prerequisites is skipped while the others continue. The result is a inference-ready layout under <LMUData root>/datasets/:

LMUData/
└── datasets/
    ├── Astro2D/
    ├── VANTAGE_2DGrounding/
    ├── VANTAGE_2DPointing/
    ├── VANTAGE_DVC/
    ├── VANTAGE_EventVerification/
    ├── VANTAGE_SOT/
    ├── VANTAGE_Temporal/
    └── VANTAGE_VQA/

Dataset Characterization

Data Collection Method
Hybrid: Human, Synthetic, Automated. Video data is sourced from vendor-provided footage (GoPro captures of warehouse and smart space environments), synthetic generation (DriveSim collision and multi-camera scenarios), and publicly scraped sources (Dubuque highway/ITS footage).

Labeling Method
Hybrid: Human, Synthetic, Pseudolabeled. Annotations for VQA, dense video captions, and temporal localization are primarily human-authored. Spatial grounding labels (2D/3D bounding boxes, referring expressions) use a combination of human annotation and pseudolabeling pipelines (detection + SAM for spatial pointing). Event verification labels are human-curated. Annotations are held server-side for evaluation only.

Evaluation

Tasks and Submission Formats

Category Task Metric
Semantic VQA Accuracy
Semantic Event Verification Macro F1
Temporal Dense Video Captioning SODA-c
Temporal Temporal Localization mIoU
Spatial 2D Object Localization F1@IoU=0.5
Spatial 2D Referring Expressions mIoU
Spatial 2D Spatial Pointing Accuracy
Spatio-Temporal Single Object Tracking AUC

See Submission Format for the expected prediction schema. Results are submitted through the official website and ranked on the Hugging Face leaderboard.

Metric Notes

  • Accuracy: Percentage of correct predictions.
  • SODA-c: Metric for dense video captioning quality across event coverage and language quality.
  • Macro F1: Unweighted mean of per-class F1 scores (harmonic mean of precision and recall).
  • F1@IoU=0.5: F1 score at an IoU threshold of 0.5.
  • mIoU: Mean Intersection over Union — average overlap between predicted and ground-truth bounding boxes (also used for temporal localization spans).
  • AUC: Area under the ROC curve, measuring the model's ability to distinguish correct detections or tracks from incorrect ones across varying confidence thresholds.

Generating Predictions

This dataset repository provides data and schemas only; it does not score predictions. Ground-truth answers are withheld and scoring happens server-side. The end-to-end workflow is:

  1. Prepare an inference-ready LMUData layout (see Quick Start):

    python scripts/run_lmudata.py --all --lmu-root ~/LMUData

  2. Run inference using VANTAGE-Bench's evaluation toolkit. Each run emits a *.submission.jsonl of predictions.

  3. Submit the predictions through the flow documented on VANTAGE-Bench's official website and in the GitHub run guides. Accepted submissions are scored server-side and ranked on the Hugging Face leaderboard.

See scripts/RUN_LMUData.md for setup, disk requirements, troubleshooting, and task-specific notes.

Submission Format

Each prediction is a single JSON record (one record per line in the submission JSONL):

{
  "id": "<task_specific_id>",
  "task": "<task_name>",
  "conversations": [
    {
      "from": "assistant",
      "value": "<raw_model_prediction>"
    }
  ],
  "metadata": {
    "model": "<model_name>",
    "extra": {}
  }
}

The authoritative, per-task submission specification lives in the GitHub repository and scripts/RUN_LMUData.md. Submit through the entry points on the official website; this repository performs no scoring or ranking.

Dataset Format

Video (mp4) and images (jpg). Only the input side of each annotation ships; ground-truth answers are withheld. For example, a VQA record carries just the question and options:

{
  "q_uid": "GX010071_Clip_4.mp4",
  "question": "How many people can exit the door at once while walking?",
  "options": ["4", "all", "3", "2"]
}

Field names vary by task (see data/README.md); no record includes an answer or ground-truth label.

Dataset Quantification

Category Task Media Entries
Semantic VQA 282 videos 1,195 QAs (MCQ)
Semantic Event Verification 163 videos 163 QAs (BCQ)
Temporal Dense Video Captioning 104 videos 717 Events
Temporal Temporal Localization 203 videos 1,067 Segments / Spans
Spatial 2D Object Localization 628 images (3 video sequences) 27,404 Bboxes
Spatial 2D Referring Expressions 1,503 images 3,276 Expressions
Spatial 2D Spatial Pointing 361 images 1,005 QAs (MCQ)
Spatio-Temporal Single Object Tracking 102 video clips 200 Trajectories

Total Entries (Annotations): 35,027 Total Media Samples (across tasks, with overlaps): 3,346 Total Data Storage: 42 GB

Disclaimers

Potential Known Risks

  • Ground truth annotations are not publicly released. All evaluation is performed server-side.
  • Some warehouse videos are concatenated clips from longer recording sessions.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

References

Citations 

@article{Sun2025RefDrone,
  author    = {Zhichao Sun and Yuda Zou and Xian Sun and Yingchao Feng and Wenhui Diao and Menglong Yan and Kun Fu},
  title     = {{RefDrone}: A Challenging Benchmark for Referring Expression Comprehension in Drone Scenes},
  journal   = {arXiv preprint arXiv:2502.00392},
  year      = {2025}
}

License/Terms of Use

This dataset is released under the NVIDIA Evaluation Data License.

Dataset Owner(s)

NVIDIA Corporation

Dataset Creation Date

April 24, 2026

Changelog

See CHANGELOG.md for release history.