README.md

metadata

license: apache-2.0
task_categories:
  - video-classification
  - question-answering
tags:
  - spatial-reasoning
  - video-qa
  - 3d-understanding
  - arkitscenes
size_categories:
  - 1K<n<10K
viewer: true

Vsi_Mcq

Dataset Description

This dataset contains 1000 video-based spatial reasoning questions from the ARKitScenes dataset. Each sample includes a video showing an indoor scene and a question about spatial relationships between objects from a specific viewpoint.

Dataset Structure

This dataset follows the VideoFolder format. The structure is:

dataset/
├── train/
│   ├── videos/
│   │   ├── video1.mp4
│   │   ├── video2.mp4
│   │   └── ...
│   └── metadata.csv
├── validation/
│   ├── videos/
│   │   ├── video1.mp4
│   │   └── ...
│   └── metadata.csv
└── README.md

Metadata Format

Each metadata.csv contains:

• file_name: Path to the video file (relative to split directory)
• original_file_name: Original filename from source dataset
• dataset_source: Source dataset identifier
• question_type: Type of spatial reasoning question
• question: The question text with detailed instructions
• options: JSON string of multiple choice options
• ground_truth: Correct answer (A, B, C, or D)

Question Types

The dataset includes the following spatial reasoning question types:

• obj_appearance_order
• object_rel_direction_easy
• object_rel_direction_hard
• object_rel_direction_medium
• object_rel_distance
• route_planning

Usage

Loading the Dataset

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("advaitgupta/vsi_mcq")

# Access train/validation splits
train_dataset = dataset['train']
val_dataset = dataset['validation']

# Example: Get first sample
sample = train_dataset[0]
print(sample['question'])
print(sample['options'])
print(sample['ground_truth'])

# The video can be accessed via sample['file_name']

Processing Videos

import cv2
import json

# Load metadata to get video info
metadata = train_dataset.to_pandas()

# Process a video
video_path = metadata.iloc[0]['file_name']
question = metadata.iloc[0]['question']
options = json.loads(metadata.iloc[0]['options'])

# Load video with OpenCV
cap = cv2.VideoCapture(video_path)
# ... your video processing code

Dataset Statistics

• Total samples: 1000
• Video format: MP4
• Source datasets: arkitscenes, scannet, scannetpp
• Question types: 6

Example Questions

Object Directional Reasoning (Medium)

"If I am standing by the stove and facing the sofa, is the tv to my left, right, or back? An object is to my back if I would have to turn at least 135 degrees in order to face it."

Object Directional Reasoning (Hard)

"If I am standing by the bathtub and facing the table, is the toilet to my front-left, front-right, back-left, or back-right? The directions refer to the quadrants of a Cartesian plane (if I am standing at the origin and facing along the positive y-axis)."

Data Quality

• All video files have been validated to exist
• Metadata is consistent across all samples
• Questions include clear instructions for spatial reasoning

Citation

If you use this dataset, please cite the original ARKitScenes paper:

@article{arkitscenes2021,
  title={3D Scene Understanding with ARKitScenes},
  author={Baruch, Gilad and others},
  journal={arXiv preprint arXiv:2111.08897},
  year={2021}
}

License

This dataset is released under the Apache 2.0 license.