README.md

---
pretty_name: STRIDE
license: mit
task_categories:
  - visual-question-answering
language:
  - en
size_categories:
  - 1K<n<10K
tags:
  - autonomous-driving
  - spatiotemporal-reasoning
  - nuscenes
  - waymo
  - vqa
configs: []
---

# STRIDE

**STRIDE: Evaluating Spatiotemporal Reasoning in Driving Edge Cases**

[Paper](https://arxiv.org/) · [Code](https://github.com/PlusLabNLP/STRIDE) · [Project page](https://pluslabnlp.github.io/STRIDE/)

This Hub repo contains **question annotations** (and nuScenes group sidecars). Raw camera images are **not** included; download them from [nuScenes](https://www.nuscenes.org/) and [Waymo Open Dataset](https://waymo.com/open/) under their terms, then follow the GitHub prepare scripts.

## Dataset summary

| Split | Questions | Source images | Files |
| ----- | --------- | ------------- | ----- |
| nuScenes | 1,150 | nuScenes trainval | `nuScenes/questions.json`, `nuScenes/formatted_metadata/` |
| Waymo | 1,200 | Waymo Open Dataset val | `Waymo/questions.json` |
| Mini | 7 (demo only) | same as above | `Mini/` |

39 templates in six families: spatial perception (SP), spatial understanding (SU), temporal memory (TM), temporal extrapolation (TE), trajectory prediction (TRJ), scene-context awareness (SC). Formats: MCQ, open-ended (OEQ), and waypoint regression (`TRJ-5` / `TRJ-6`). Each question uses five frames.

## Load

```python
from huggingface_hub import hf_hub_download
import json

path = hf_hub_download(
    repo_id="uclanlp/STRIDE",
    filename="nuScenes/questions.json",
    repo_type="dataset",
)
payload = json.load(open(path))
tasks = payload["tasks"]
print(payload["meta"]["n_questions"], tasks[0]["id"], tasks[0]["question"][:80])
```

For inference and scoring, clone the [code repo](https://github.com/PlusLabNLP/STRIDE) and run `scripts/prepare_*_media.py` / `python -m stride.cli score`.

## License

STRIDE annotation files in this repository are released under **MIT**. nuScenes and Waymo images remain under their original licenses and must be obtained from those datasets. Do not redistribute those images through this Hub repo.