---
license: other
task_categories:
- visual-question-answering
language:
- en
tags:
- autonomous-driving
- driving-vqa
- waymo
- video-qa
- safety-critical
pretty_name: WaymoQA VideoQA (test subset)
---

# WaymoQA — VideoQA test subset (mosaic frames)

This repository hosts the **video** portion of the **WaymoQA** *test* split, prepared for
VideoQA evaluation. It contains the multi-view **3x3 mosaic frames** for every Waymo
scenario token that carries video questions in the test set.

## Contents

| File | Description |
|------|-------------|
| `mosaics.tar.part_aa` … `mosaics.tar.part_ah` | Split archive (8 × 2 GiB) of the mosaic frames |
| `mosaics.tar.part_ai` | Final split of the archive |
| `test.jsonl` | Full WaymoQA test QA annotations (contains both image and video questions) |

## What the data is

- **Source dataset:** [Waymo Open Dataset — End-to-End Driving](https://waymo.com/open/download/)
  (`waymo_open_dataset_end_to_end_camera_v_1_0_0`, test split).
- **QA annotations:** the official **WaymoQA** dataset (Yu et al., arXiv:2511.20022),
  downloaded from the authors' release.
- **Scope:** the **video** questions of the WaymoQA test split —
  **896 QA items** over **54 unique scenario tokens**.
- **Frames:** **7,357** mosaic images (`{token}_{frame_index:03d}.jpg`).

### Frame format

Each `.jpg` is a **3×3 mosaic** combining the 8 synchronized E2E camera views of one
timestep, with per-camera text labels, matching the official WaymoQA video-QA
preprocessing:

```
[FRONT_LEFT] [FRONT] [FRONT_RIGHT]
[SIDE_LEFT]  (blank) [SIDE_RIGHT]
[REAR_LEFT]  [REAR]  [REAR_RIGHT]
```

Frames of a given token form a temporal sequence (the video). A token's clip is ~10 Hz.

### Note on frame counts

Each video contains ~121–141 frames (the official per-token annotation lists ~147–150).
The first few frames of each clip are **not present in the public Waymo E2E test shards**,
so only the frames actually available in the source data are included here. This does not
affect strided sampling used for evaluation.

## How to use

Merge the split archive and extract:

```bash
cat mosaics.tar.part_* > mosaics.tar
tar -xf mosaics.tar -d mosaics
```

This yields a flat directory `mosaics/` of JPEGs named `{token}_{frame:03d}.jpg`.

Select the video questions from the annotations:

```python
import json
rows = [json.loads(l) for l in open("test.jsonl", encoding="utf-8")]
video_qa = [r for r in rows if r.get("type") == "video"]   # 896 items
# each item: {token, waymo_split, type, question, category, options[4], answer_index}
```

To build a video-QA input, group the mosaic frames by `token` (sorted by frame index)
and pass them (sampled by a stride) together with the `question` / `options`.

## Attribution & license

- Waymo Open Dataset is released under the
  [Waymo Dataset License Agreement for Non-Commercial Use](https://waymo.com/open/terms/).
  Raw sensor data is **not** redistributed here; only derived mosaic frames of a small
  subset are provided, following the original Waymo terms.
- The QA annotations belong to the **WaymoQA** authors:
  Yu, Seungjun; Lee, Seonho; Kim, Namho; Shin, Jaeyo; Park, Junsung; Ryu, Wonjeong;
  Jung, Raehyuk; Shim, Hyunjung.
  *WaymoQA: A Multi-View Visual Question Answering Dataset for Safety-Critical Reasoning
  in Autonomous Driving.* arXiv:2511.20022.
- Please cite the original WaymoQA paper and comply with the Waymo dataset terms.

## Citation

```bibtex
@article{yu2025waymoqa,
  title={WaymoQA: A Multi-View Visual Question Answering Dataset for Safety-Critical Reasoning in Autonomous Driving},
  author={Yu, Seungjun and Lee, Seonho and Kim, Namho and Shin, Jaeyo and Park, Junsung and Ryu, Wonjeong and Jung, Raehyuk and Shim, Hyunjung},
  journal={arXiv preprint arXiv:2511.20022},
  year={2025}
}
```