E3VQA

E3VQA is the benchmark introduced in:

Towards Comprehensive Scene Understanding: Integrating First and Third-Person Views for LVLMs
Insu Lee^*, Wooje Park^*, Jaeyun Jang, Minyoung Noh, Kyuhong Shim^†, and Byonghyo Shim^† (* Equal contribution, † Corresponding authors)
NeurIPS 2025 Spotlight

Paper | Code

E3VQA is a multi-view visual question answering benchmark designed to evaluate scene understanding across synchronized egocentric and exocentric views.

The current dataset release contains 4,500 multiple-choice question-answer pairs from two data sources:

• Ego-Exo4D: 4,000 samples
• LEMMA: 500 samples

Each sample contains an egocentric image, an exocentric image, a question, four answer choices, and the ground-truth answer.

Categories

E3VQA contains four evaluation categories:

• Action
• Object
• Numerical
• Spatial

Each question is additionally associated with a target perspective:

• ego
• exo

Dataset Structure

Each example contains the following fields:

• id: unique sample identifier
• source: source dataset (Ego-Exo4D or LEMMA)
• category: question category
• perspective: target perspective (ego or exo)
• ego: egocentric image
• exo: exocentric image
• question: question text
• options: four multiple-choice options
• answer: ground-truth answer
• answer_ego: answer obtainable from the egocentric view
• answer_exo: answer obtainable from the exocentric view

Usage

from datasets import load_dataset

dataset = load_dataset("SNU-ISLAB/E3VQA")

sample = dataset["test"][0]

print(sample["question"])
print(sample["options"])
print(sample["answer"])

ego_image = sample["ego"]
exo_image = sample["exo"]

Evaluation

E3VQA is evaluated using multiple-choice accuracy.

Performance can be reported at multiple levels:

Evaluation axis	Subsets
Overall	All samples
Source	Ego-Exo4D, LEMMA
Category	Action, Object, Numerical, Spatial
Perspective	Ego, Exo

Standardized evaluation support for lmms-eval and VLMEvalKit will be provided.

Citation

If you use E3VQA in your research, please cite:

@inproceedings{lee2025towards,
 author = {Lee, Insu and Park, Wooje and Jang, Jaeyun and Noh, Minyoung and Shim, Kyuhong and Shim, Byonghyo},
 booktitle = {Advances in Neural Information Processing Systems},
 doi = {10.52202/085713-0627},
 pages = {18576--18628},
 title = {Towards Comprehensive Scene Understanding: Integrating First and Third-Person Views for LVLMs},
 url = {https://proceedings.neurips.cc/paper_files/paper/2025/file/1af83ab66b4f07a3f55788e67dab5782-Paper-Conference.pdf},
 volume = {38, Main Conference},
 year = {2025}
}