gatilin/argus-visual-decision

ARGUS Visual Decision Dataset

A large-scale visual decision-making dataset for training and evaluating Vision-Language Models (VLMs).

Overview

• Total samples: 16,000
• Tasks: 16 visual decision primitives
• Language: Chinese (zh)
• Format: Parquet with images

Tasks

• color_recognition: 1000
• danger_level: 1000
• defect_detection: 1000
• existence_check: 1000
• hazard_detection: 1000
• indoor_outdoor: 1000
• object_category_coarse: 1000
• object_counting_person: 1000
• object_counting_vehicle: 1000
• ood_detection: 1000
• road_condition: 1000
• scene_classification: 1000
• spatial_relation: 1000
• texture_material: 1000
• time_of_day: 1000
• weather_recognition: 1000

Primitives

Primitive	Description	Example
VisChoice	Visual multiple choice	"What color is the object?" → "red"
VisScore	Visual scoring/rating	"Rate the damage level" → "3"
VisCount	Visual counting	"How many people?" → "5"
VisExist	Existence verification	"Is there a defect?" → "yes"

Usage

from datasets import load_dataset

ds = load_dataset("linxu/argus-visual-decision")
print(ds["train"][0])

Dataset Structure

├── data/
│   ├── train-00000-of-00001.parquet
│   └── test-00000-of-00001.parquet
└── images/
    └── <source_dataset>/<split>/<image_files>

Source Datasets

Images are sourced from public benchmarks: COCO2017, DeepFashion2, BDD100K, DOTA, Objects365, StreetHazards, ImageNet, and others.

Citation

@dataset{argus-visual-decision,
  title={ARGUS Visual Decision Dataset},
  author={Lin Xu et al.},
  year={2026}
}