zhouhanjing/LogiScope-VQA-preview

Dataset Card for LogiScope-VQA (Preview)

⚠️ This is a preview release. It contains approximately 10% of the full LogiScope-VQA benchmark, provided for academic review during the submission period. The complete version (10,274 VQA pairs) will be publicly released after paper acceptance.

Dataset Description

LogiScope-VQA is the first dedicated benchmark for evaluating Large Multimodal Models (LMMs) on logistics hazard identification in real-world industrial scenarios. It is built from 3.5 million surveillance clips collected over one year across the global warehouse parks of Cainiao, a leading intelligent logistics company.

The full benchmark provides:

• Visual samples: 5,394 in total — 2,476 images and 2,918 video clips, primarily from private warehouse surveillance, complemented by open-source and synthetic data (covering high-rack areas, loading docks, storage zones, picking shelves, security screening areas, day/night scenes, etc.).
• VQA pairs: 10,274 human-validated questions in two formats:
  • Multiple-choice: each question has 2–16 options, including a distractor option ("the queried object does not exist") to discourage hallucinated answers.
  • Open-ended: free-form safety analysis questions (e.g., listing all violations in a scene with consequence analysis); options is empty and answer is a reference analysis in text.
• Coverage: 18 core objects and 20 risk types grounded in real-world warehouse safety regulations.
• Task taxonomy: 3 competency dimensions → 10 tasks → 39 subtasks, organized as a progressive curriculum:
  1. Industrial Element Perception — fine/coarse-grained perception of dense, low-resolution surveillance targets
  2. Warehouse Knowledge Understanding — commonsense, spatial relations, and operator roles
  3. Potential Risk Reasoning — perimeter access control, fire monitoring, personnel safety duty, and equipment operation compliance

Note: the statistics above describe the full benchmark. This preview repository contains a 10% subset:

	Full benchmark	This preview
VQA pairs	10,274	1,000 (950 multiple-choice, 50 open-ended)
Visual samples	5,394 (2,476 images, 2,918 video clips)	498 (177 images, 321 video clips)

Preview Subset

This repository hosts a preview subset of 1,000 VQA pairs (~10% of the full benchmark) — 950 multiple-choice and 50 open-ended questions — built on 498 unique visual samples (177 images and 321 video clips), sampled to preserve the task and category coverage of the complete release:

• Identical data format, fields, and annotation standards as the full version.
• Intended for academic evaluation and reproduction during the review period.
• Results on this subset should not be reported as results on the full benchmark.

Dataset Structure

Data Fields

Field	Type	Description
index	int	Sample ID
question	string	Question text (Chinese)
question_type	string	"open-ended" for open-ended questions; "option" for multiple-choice questions
options	list[string]	Option list for multiple-choice questions (e.g., A. Yes); empty for open-ended
answer	list[string] / string	Multiple-choice: correct option letter(s), e.g., ["B"]; open-ended: reference analysis text
data_type	string	images / videos
images	list[string]	Relative paths to image files
videos	list[list[string]]	Each video is an ordered list of frame paths (JPEGs) under a per-video directory

All media paths are relative to the repository root.

Repository Layout

├── README.md
├── test.json                # QA annotations (preview subset)
└── test/
    ├── img_*.jpg            # images (flat layout)
    └── vid_*/               # one directory per video
        ├── vid_*_f00.jpg    # ordered frames
        └── ...

Evaluation Metric

• Multiple-choice: accuracy — the predicted option letter must exactly match answer.
• Open-ended: scored by comparing the model output against the reference answer, which has been human-validated.

Ethics

For the private data, all privacy-sensitive content — clearly visible human faces, warehouse names, and commercial identifiers — has been blurred or masked via automated detection with manual verification; synthetic and open-source data are released as-is, as they contain no real-world identities. The dataset contains no data collected in the United States, and its construction and release comply with relevant ethical guidelines and data protection regulations. This dataset is intended solely for evaluating and researching model safety capabilities, and must not be used for surveillance of real individuals or any discriminatory purposes.

Citation

If you find LogiScope-VQA useful in your research, please cite:

@article{zhou2026logiscope-vqa,
  title   = {LogiScope-VQA: Benchmarking Vision-Language Models for Logistics Hazard Identification in Industrial Scenarios},
  author  = {Hanjing Zhou, Mingze Yin, Ying Lian, Jun Ma, Chang-Yu Hsieh, Yanbing Zhou},
  journal = {arXiv preprint arXiv:2609.09790},
  year    = {2026}
}