Crash-VQA

Dataset, model, and tools for multi-view post-crash analysis.

48,871 Task instances

97,323 Post-crash images

12,734 Vehicles

9,913 Crash cases

This repository provides the lightweight tools for browsing Crash-VQA, running CrashQwen-8B on multi-view image packs, evaluating predictions, and fine-tuning on the public dataset.

Tasks

Task	Prediction
plane_atomic	impact plane: front / rear / left / right
clock_atomic	principal direction of force: clock position 1–12
extent_atomic	deformation extent: minor / moderate / severe
deltav_atomic	total Delta-V: 0–10 / 10–20 / 20–30 / 30+ km/h
ais2_atomic	whether any occupant sustained AIS 2+ injury

Up to nine images of the same vehicle are used in canonical order: front, rear, right, left, front-right, front-left, rear-right, rear-left, top.

Example case

Representative full-case visualization with the nine canonical views, available vehicle metadata, and all five Crash-VQA targets.

The raw views used for inference are stored in examples/example_case/images/ under their canonical filenames. The preview above is a compact visualization of the same case.

Install

git clone https://github.com/oValach/Crash-VQA.git
cd Crash-VQA
pip install -e ".[inference]"

For 4-bit inference:

pip install -e ".[inference,quant]"

If access to the Hugging Face repositories requires authentication:

hf auth login

Browse the dataset

Open a local multi-view browser:

python scripts/browse.py --config natural --split test

Then open http://127.0.0.1:8000.

Run CrashQwen

Run the bundled nine-view example:

python scripts/infer.py examples/example_case --task plane_atomic --load-in-4bit

Run all five tasks and optionally provide vehicle metadata:

python scripts/infer.py examples/example_case \
  --task all \
  --vehicle-type "Sedan/Saloon" \
  --model-year 2013 \
  --curb-wt-kg 1307 \
  --load-in-4bit

For your own case, use canonical filenames. Missing views can simply be omitted:

my_case/
└── images/
    ├── front.jpg
    ├── rear.jpg
    ├── right.jpg
    ├── left.jpg
    ├── front_right.jpg
    ├── front_left.jpg
    ├── rear_right.jpg
    ├── rear_left.jpg
    └── top.jpg

Vehicle metadata is optional. CrashQwen uses vehicle_type, model_year, and curb_wt_kg for Delta-V and AIS2+ when these values are available.

Evaluate

python scripts/evaluate.py \
  --config natural \
  --split test \
  --load-in-4bit \
  --output outputs/crashqwen.jsonl

Use --limit 100 for a quick run. Predictions are written to JSONL and metrics to outputs/crashqwen.metrics.json.

Fine-tune

Install the training dependencies:

pip install -e ".[train]"

Run the reference fine-tuning script on the public training split:

python scripts/train.py --output-dir outputs/crashqwen-train

Repository layout

Crash-VQA/
├── README.md
├── LICENSE
├── pyproject.toml
├── .gitignore
├── examples/
│   └── example_case/
│       ├── preview.jpg
│       └── images/
│           ├── front.jpg
│           ├── front_right.jpg
│           ├── right.jpg
│           ├── rear_right.jpg
│           ├── rear.jpg
│           ├── rear_left.jpg
│           ├── left.jpg
│           ├── front_left.jpg
│           └── top.jpg
├── scripts/
│   ├── infer.py
│   ├── evaluate.py
│   ├── train.py
│   └── browse.py
└── src/
    └── crashvqa/
        ├── __init__.py
        ├── data.py
        ├── tasks.py
        ├── model.py
        └── metrics.py

License

Repository code is licensed under Apache-2.0. Crash-VQA is distributed separately under CC BY-NC 4.0 and is derived from NHTSA crash-investigation records. See the Hugging Face dataset page for dataset terms and provenance.

Citation

@inproceedings{valach2026crashvqa,
  title     = {Crash-VQA: A Multi-View Dataset for Post-Crash Mechanics and Severity Prediction},
  author    = {Valach, Ondřej and Gruber, Ivan and Diviš, Václav},
  booktitle = {ECCV 2026 MARS2 Workshop},
  year      = {2026}
}