Sign in to view source links and access this dataset
Description
203,048 question-answer pairs accompany 2,507 video trajectories generated from 1,261 procedurally created indoor scenes. The dataset was created by Ellis Brown, Arijit Ray, Ranjay Krishna, Ross Girshick, Rob Fergus, and Saining Xie for training models in spatial video understanding. It was last updated in November 2025.
Use Cases
Train visual-language models to answer spatial questions about video trajectories.
Benchmark model performance on video-based question-answer pairs across synthetic indoor scenes.
Fine-tune models for spatial reasoning using procedurally generated video and text data.
Strengths
203,048 question-answer pairs provide a substantial volume of training examples.
Data is derived from 2,507 video trajectories across 1,261 distinct synthetic scenes.
Limitations
Dataset is entirely synthetic, which may not reflect real-world visual complexity and distribution.
Specific column names, data formats, and sample sizes are not provided in the input.
Provenance
Source
huggingface
Collection Method
Procedurally generated from synthetic indoor scenes.