Datasets:

Litian2002
/

spatialvlm_qa

license: mit
task_categories:
  - visual-question-answering
language:
  - zh
  - en
tags:
  - spatial-reasoning
  - blender
  - synthetic
  - vision-language
pretty_name: Synthetic Spatial Visual Language Question Answering Dataset
size_categories:
  - 1K<n<100K

Synthetic Spatial Visual Language Question Answering Dataset

Automatically generated from the Blender Scene Dataset. Each example contains an image and a question-answer pair to probe metric (numeric) and relation (true/false) spatial reasoning skills.

Images: Rendered using Blender (1000 scenes, 5 random primitives each, random cameras and lighting).
Metadata: Object name, position, scale, color, material flags.
Questions: 10 per image, drawn from handcrafted templates, covering:
Euclidean/horizontal/vertical distance queries
Object width queries
Front/back predicates relative to the camera

Distances are in Blender scene units, rounded to 1 cm precision.

Fields

Field	Type	Description
image	image	rendered scene
question	string	natural language query
answer	string	ground truth answer

Quotes

@misc{synthetic_spatialvlm_qa,
title = {Synthetic Spatial Vision-Language Question Answering Dataset},
author = {<your name>},
year = 2025,
url = {https://huggingface.co/datasets/Litian2002/spatialvlm_qa}
}