SpaceVista/Guinness_World_Records

Guinness World Records – Abnormal Size VQA

A curated visual question-answering dataset featuring objects of extraordinary or record-breaking dimensions, sourced from Guinness World Records. Each image is paired with one or more measurement questions (height, length, width, diameter) and ground-truth numeric answers.

---

Dataset Preview

Below is a sample of entries from the dataset. Images are stored under images/ and referenced by filename.

#	Image	Object	Question	Answer
1		rifle	What is the length of this rifle (in meters)?	10.18 m
2		telephone	How tall is this telephone (in meters)?	2.47 m
4		monster truck	What is the length of this monster truck (in meters)?	9.75 m
5		hiking boot	How tall is this hiking boot (in meters)?	4.20 m
11		man-made structure	How tall is this man-made structure (in meters)?	828 m
15		bicycle	How long is this bicycle (in meters)?	55.16 m
18		Axleless Ferris Wheel	How tall is this Axleless Ferris Wheel (in meters)?	142.52 m
38		motorcycle	How long is this motorcycle (in meters)?	26.29 m

---

Dataset Summary

This dataset is designed to benchmark the physical size and dimension estimation capabilities of vision-language models (VLMs). It focuses on objects that hold or are associated with Guinness World Records for their abnormal dimensions — covering outdoor scenes, indoor scenes, drone aerial views, and tabletop/close-up shots.

Split	Images	QA Pairs
all	33	50

Scene Label Distribution

Scene Label	# QA Pairs
outdoor	35
indoor	9
drone	4
tabletop	2

Measurement Types

All questions ask for a single numeric measurement in meters or centimeters:

• Height (How tall is …)
• Length (How long is … / What is the length of …)
• Width (How wide is … / What is the width of …)
• Diameter (What is the diameter of …)

---

File Structure

.
├── images/
│   ├── 1.jpg          # world's longest rifle
│   ├── 2.jpg          # world's tallest telephone
│   ├── 3.jpg          # world's longest monster truck
│   └── ...            # 30 more images (1–33)
├── annotation.csv     # flat CSV with Num., Image_Name, Question, Answer, Object, Scene
└── annotation.jsonl   # structured JSONL with full metadata per QA pair

annotation.csv columns

Column	Description
Num.	QA pair index (1–50)
Image_Name	Filename of the corresponding image
Question	Natural-language measurement question
Answer	Ground-truth numeric answer
Object	The object being measured
indoor/outdoor/tiny/tabletop/drone	Scene type label

annotation.jsonl fields

Field	Description
Question Number	QA pair index
Question	Natural-language measurement question
Answer	Ground-truth numeric value
Scene Source	Internal collection identifier
Frames Path	Image filename (without extension)
Input Type	Image+text for all entries
Input Point / Mask / BBox	Optional spatial prompt (null for this subset)
Question Type	regression (numeric answer)
TaskType	HEIGHT for all entries
Object	The record-holding object
Scene Label	outdoor / indoor / drone / tabletop

---

Usage

from datasets import load_dataset

ds = load_dataset("SpaceVista/Guinness_World_Records")

Or load annotations manually:

import json
from pathlib import Path
from PIL import Image

data = [json.loads(l) for l in open("annotation.jsonl")]

for entry in data:
    img_path = Path("images") / f"{entry['Frames Path']}.jpg"
    image = Image.open(img_path)
    print(entry["Question"], "→", entry["Answer"])

---

Data Provenance and Licensing

⚠️ Important Notice Regarding Data Licensing

This dataset contains images collected from multiple sources:

• A subset of the images were obtained directly through the official Guinness World Records website (guinnessworldrecords.com). These images are identified by their association with official record entries and were retrieved via publicly accessible web pages.

• The remaining images were gathered from various third-party web sources (news articles, social media, photo repositories, etc.) during data collection. We suspect that these images may carry potential licensing restrictions or unresolved copyright claims. We were unable to verify the original license terms for each of these images at collection time.

We strongly advise users of this dataset to:

1. Treat all images as potentially copyrighted unless verified otherwise.
2. Use this dataset for non-commercial research and evaluation purposes only.
3. Not redistribute individual images outside the context of this dataset.
4. Independently verify licensing before any commercial or production use.

The annotations (questions and numeric answers) were produced by the dataset authors and are released for research use.

---

Citation

If you use this dataset in your research, please cite it as:

@article{sun2025spacevista,
  title={SpaceVista: All-Scale Visual Spatial Reasoning from mm to km}, 
  author={Sun, Peiwen and Lang, Shiqiang and Wu, Dongming and Ding, Yi and Feng, Kaituo and Liu, Huadai and Ye, Zhen and Liu, Rui and Liu, Yun-Hui and Wang, Jianan and Yue, Xiangyu},
  journal={arXiv preprint arXiv:2510.09606},
  year={2025}
}