README.md

metadata

license: cc-by-nc-4.0
language:
  - en
pretty_name: MVVBench
size_categories:
  - 10K<n<100K
task_categories:
  - video-classification
  - visual-question-answering
  - video-text-to-text
tags:
  - video
  - annotations
  - ego-exo4d
  - cmu-panoptic
  - mmptrack
  - multi-view
source_datasets:
  - extended|other-ego-exo4d
  - extended|other-cmu-panoptic
  - extended|other-mmptrack
annotations_creators:
  - expert-generated
language_creators:
  - found
multilinguality:
  - monolingual
extra_gated_prompt: >
  This dataset contains only annotations. Users must separately obtain the
  source videos from Ego-Exo4D, CMU Panoptic Studio, and MMPTRACK under each
  dataset's original license. Commercial use is restricted due to source-dataset
  terms.
extra_gated_fields:
  Name: text
  Affiliation: text
  I agree to use this dataset for non-commercial research only: checkbox
  I have read and will comply with the source dataset licenses: checkbox

Multi-View Video QA Dataset

This dataset contains 1,323 question-answer pairs across 608 multi-view video samples for video question answering research.

Repository Structure

.
├── annotations/
│   ├── README.md                 # This file
│   ├── all_questions.json        # QA annotations (original format)
│   ├── all_questions.parquet     # QA annotations (parquet format)
│   ├── video_mapping.csv         # Video file mapping (source -> target)
│   └── prepare_videos.py         # Script to prepare videos from source datasets
│
├── dataset/                      # Source datasets (user must download)
│   ├── egoexo/
│   │   └── takes/
│   │       ├── sfu_basketball_01_10/
│   │       │   └── frame_aligned_videos/
│   │       │       └── downscaled/
│   │       │           └── 448/
│   │       │               ├── aria01_1201-1.mp4
│   │       │               ├── cam01.mp4
│   │       │               └── ...
│   │       └── ...
│   │
│   ├── mmptrack/
│   │   ├── train/
│   │   │   └── videos/
│   │   │       ├── 63am/
│   │   │       │   └── cafe_shop_0/
│   │   │       │       ├── cafe_shop_0_camera1.mp4
│   │   │       │       └── ...
│   │   │       └── 64am/
│   │   └── validation/
│   │       └── videos/
│   │           └── 64pm/
│   │
│   └── panoptic/
│       ├── 160224_haggling1/
│       │   └── hdVideos_down/
│       │       ├── hd_00_00.mp4
│       │       └── ...
│       └── ...
│
└── videos/                       # Prepared videos (output of prepare_videos.py)
    ├── sfu_basketball_05_35_view0.mp4
    ├── sfu_basketball_05_35_view1.mp4
    ├── 63am_cafe_shop_0_view0.mp4
    ├── 160224_haggling1_view0.mp4
    └── ...

Annotation Files

all_questions.json / all_questions.parquet

Column	Description
sample_name	Video sample identifier
views	List of view indices used for the question (e.g., [3, 2])
question	Question text
choice_A, choice_B, choice_C, choice_D	Multiple choice options
correct_answer	Correct answer key (A/B/C/D)
gt_answer_text	Ground truth answer text
category	Question category
static_or_dynamic	Static or dynamic scene
timestamp_start, timestamp_end	Relevant time range (seconds)
question_type	Type: counting, descriptive, binary
original_negated	Original or negated question

video_mapping.csv

Maps 3,088 video files from source datasets to standardized names:

Column	Description
target	Target filename (e.g., sfu_basketball_05_35_view0.mp4)
sample_name	Sample name
view	View number
dataset	Source dataset: egoexo, mmptrack, panoptic
split	For mmptrack: train or validation
source_rel_path	Relative path within source dataset
source_file	Source filename

Preparing Videos

Due to licensing restrictions, videos are not included. You must download them from the original sources and use the provided script to prepare them.

Step 1: Download Source Datasets

1. Ego-Exo4D (2,590 files)

   • Website: https://ego-exo4d-data.org/
   • Download the takes with frame_aligned_videos/downscaled/448/ videos

2. MMPTrack (272 files)

   • Website: https://iccv2021-mmp.github.io/subpage/dataset.html
   • Download both train and validation splits

3. CMU Panoptic (220 files)

   • Website: http://domedb.perception.cs.cmu.edu/
   • Download sequences with hdVideos_down/ (downscaled HD videos)

Step 2: Organize Directory Structure

Place downloaded datasets under dataset/ directory following the structure shown above.

Step 3: Run the Preparation Script

# Preview first
python annotations/prepare_videos.py \
    --dataset-root ./dataset \
    --output-dir ./videos \
    --dry-run

# Copy files
python annotations/prepare_videos.py \
    --dataset-root ./dataset \
    --output-dir ./videos

# Or create symlinks (saves disk space)
python annotations/prepare_videos.py \
    --dataset-root ./dataset \
    --output-dir ./videos \
    --symlink

Statistics

Dataset	Samples	Video Files	Categories
Ego-Exo4D	496	2,590	Basketball, cooking, bouldering, dance, music, etc.
MMPTrack	68	272	Cafe, retail, office, lobby, industry safety
CMU Panoptic	44	220	Haggling, mafia, ultimatum, social games
Total	608	3,082

Citation

If you use this dataset, please cite the original source datasets:

@article{grauman2024egoexo4d,
  title={Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives},
  author={Grauman, Kristen and others},
  journal={CVPR},
  year={2024}
}

@inproceedings{han2021mmptrack,
  title={MMPTrack: Large-scale Densely Annotated Multi-camera Multiple People Tracking Benchmark},
  author={Han, Xiaotian and others},
  booktitle={ICCV Workshop},
  year={2021}
}

@inproceedings{joo2015panoptic,
  title={Panoptic Studio: A Massively Multiview System for Social Motion Capture},
  author={Joo, Hanbyul and others},
  booktitle={ICCV},
  year={2015}
}