Crash-VQA

Crash-VQA: Visual Question Answering for Real-World Crash Understanding

Nine canonical views of one damaged vehicle. Missing views are represented as null.

Overview

Crash-VQA is a multi-image benchmark for structured post-crash vehicle interpretation. Each row contains one task annotation, a normalized label, structured vehicle metadata, and up to nine canonical views of the same vehicle. The release supports both natural-distribution and balanced configurations for training and evaluation.

At a glance

Metric	Value
QA rows	64,487
Unique images	97,323
Unique image packs	13,782
Unique cases	9,913
Unique vehicles	12,734
Tasks	5
Canonical image views	9

What makes Crash-VQA useful

• Multi-view evidence: each sample can expose front, rear, side, corner, and top views independently.
• Structured targets: all five tasks use compact normalized label spaces suitable for reproducible evaluation.
• Two data regimes: natural preserves the observed distribution, while balanced supports controlled comparisons.
• Vehicle context: optional vehicle type, model year, and curb weight are available as structured metadata.
• Case-isolated splits: public identifiers support analysis without exposing source filesystem paths.

Example task rows

The montage shows one representative image for each task. A model receives the available multi-view image set and predicts the task-specific label.

Quick start

from datasets import load_dataset

natural = load_dataset("oValach/Crash-VQA", "natural")
balanced = load_dataset("oValach/Crash-VQA", "balanced")

row = natural["train"][0]
print(row["task"], row["label"], row["metadata"])

available_images = {
    name: row[name]
    for name in ['image_front', 'image_rear', 'image_right', 'image_left', 'image_front_right', 'image_front_left', 'image_rear_right', 'image_rear_left', 'image_top']
    if row[name] is not None
}

Explore the dataset

• Natural / train
• Natural / validation
• Natural / test
• Balanced / train
• Balanced / test
• Data Studio

Tasks

Task	Prediction target	Allowed labels	Rows
plane_atomic	Principal impact plane	front, rear, left, right	15,316
clock_atomic	Impact direction on a clock face	1 through 12	15,650
extent_atomic	Damage extent	minor, moderate, severe	15,506
deltav_atomic	Delta-V interval	0-10, 10-20, 20-30, 30+	11,897
ais2_atomic	AIS 2+ injury indicator	true, false	6,118

Configurations and splits

Configuration	Split	Rows
natural	train	36,632
natural	validation	4,902
natural	test	7,337
balanced	train	13,202
balanced	test	2,414

• natural is the default configuration and contains train, validation, and test splits.
• balanced contains train and test splits and is intended for controlled task/label comparisons.

Dataset structure

The public representation is deliberately compact: one row equals one task annotation.

Column	Type	Description
case_id	string	Stable public case identifier
split	string	Train, validation, or test
task	string	One of the five task names
label	string	Normalized task label
metadata	struct	Vehicle type, model year, and curb weight
image_front	Image() or null	Front vehicle view
image_rear	Image() or null	Rear vehicle view
image_right	Image() or null	Right vehicle view
image_left	Image() or null	Left vehicle view
image_front_right	Image() or null	Front-right vehicle view
image_front_left	Image() or null	Front-left vehicle view
image_rear_right	Image() or null	Rear-right vehicle view
image_rear_left	Image() or null	Rear-left vehicle view
image_top	Image() or null	Top vehicle view
vehicle_id	string	Stable public vehicle identifier
sample_id	string	Stable QA-row identifier

The nine image columns are Hugging Face Image() features. Path-backed cells use repository-relative paths; the raw Parquet representation may contain a nullable bytes member, while the Dataset Viewer renders the image itself.

Image views

Column	Canonical view
image_front	Front
image_rear	Rear
image_right	Right
image_left	Left
image_front_right	Front-right
image_front_left	Front-left
image_rear_right	Rear-right
image_rear_left	Rear-left
image_top	Top

Vehicle metadata

Field	Type	Meaning
vehicle_type	string or null	Source vehicle body/type description
model_year	int64 or null	Vehicle model year
curb_wt_kg	float64 or null	Curb weight in kilograms

Nulls represent genuinely unavailable source values.

Source and provenance

The source imagery and vehicle/crash metadata originate from the National Highway Traffic Safety Administration Crash Investigation Sampling System (CISS). CISS collects detailed information from a representative sample of crashes to support vehicle-safety research. Crash-VQA adds the public task formulation, normalized labels, split definitions, identifiers, packaging, and documentation.

See SOURCE_NOTICE.md for source attribution, source terms, and the non-endorsement notice.

Intended uses

Crash-VQA is intended for:

• multi-image visual question answering research;
• post-crash vehicle damage understanding;
• structured crash-analysis benchmarking;
• evaluation of multimodal models across image and vehicle-metadata inputs;
• controlled comparison of natural and balanced task distributions.

Out-of-scope uses

Crash-VQA should not be used as the sole basis for:

• medical, legal, insurance, or accident-liability decisions;
• identifying people, owners, or specific crash locations;
• operational vehicle-safety certification;
• claims that exceed the observable evidence or supplied metadata.

Limitations and responsible use

The dataset may reflect sampling, reporting, geography, vehicle-fleet, crash-severity, image-quality, and missing-metadata biases from the source collection. Damage appearance may be ambiguous across views, and labels simplify complex crash phenomena into discrete benchmark targets. Users should report performance separately by task and should not interpret benchmark accuracy as real-world accident-reconstruction competence.

License

Crash-VQA original annotations, normalized labels, task definitions, split definitions, code, and documentation are licensed under CC BY 4.0 as described in LICENSE. Source materials are not relicensed by this repository; their attribution and applicable source notices are documented separately in SOURCE_NOTICE.md.

Paper

The accompanying manuscript is titled “Crash-VQA: Visual Question Answering for Real-World Crash Understanding.”

The current intended venue is the MARS² Workshop at ECCV 2026. This wording indicates a planned submission only and must not be changed to “accepted,” “published,” or “presented” unless that status is confirmed.

Citation

A dedicated Crash-VQA citation will be added after the manuscript has an official public record. Until then, cite this dataset repository and include the manuscript title above. Any existing citation file must contain only Crash-VQA-specific or source-specific entries and must not cite unrelated datasets as the primary Crash-VQA reference.