OccStress/OccStress

---
license: other
pretty_name: OccStress
tags:
- autonomous-driving
- occupancy-prediction
- 3d-occupancy
- robustness
- nuscenes
- croissant
size_categories:
- 100K<n<1M
task_categories:
- other
---

# OccStress

OccStress is an evaluation-only robustness benchmark for 3D/4D occupancy prediction on nuScenes-style occupancy data. It is designed to measure how downstream occupancy forecasting systems respond to stress conditions in the occupancy state space and in upstream perception outputs.

This is an anonymous release candidate for review. The dataset repository is `https://huggingface.co/datasets/OccStress/OccStress`; keep the citation fields anonymous for double-blind review and replace them before public release.

## Dataset Status

- Version: `0.1.0-rc1`
- Split: validation/evaluation only
- Source payload files: `432462`
- Source payload bytes: `104537523547`
- HF staging files: `264`
- HF staging bytes: `107003337267`
- Shards: `12` tar files
- Created: `2026-05-06`
- Reviewer sample subset: included as `sample/OccStress_mini.tar` (`610` anchor samples, `10` per minimal category). See `sample/README.md` for the sample construction procedure.

## Contents

`OccStress_HF` uses a shard-based layout to avoid uploading hundreds of thousands of small files directly:

```text
OccStress_HF/
README.md
LICENSE
CITATION.cff
dataset_infos.json
croissant.json
events/ # grouped JSONL event metadata
manifests/ # file manifests, checksums, shard member lists, audits
meta/ # protocol and corruption summaries
protocols/ # evaluation protocol pickles
sample/ # reviewer mini sample tarball and checksums
shards/ # tar shards containing occ/ and cache/ payloads
viewer/ # lightweight static HTML visualization utility
```

The high-cardinality `occ/` and `cache/` trees are stored only inside `shards/*.tar`. The raw small-file directories are intentionally absent from this HF layout.

## Reviewer Viewer

The release includes a lightweight static viewer that can visualize either the
mini archive or the full sharded release without running model evaluation:

```bash
python viewer/visualize_occstress.py \
--input sample/OccStress_mini.tar \
--output occstress_viewer
```

This safely extracts the mini archive, samples protocol records, renders the
referenced occupancy states as bird's-eye-view sheets, and writes
`occstress_viewer/review.html`. The same script also accepts an extracted
`mini/` directory or the full release root; for the full release it extracts
only the small number of needed files from `shards/*.tar`. See
`viewer/README.md` for options.

## Benchmark Tracks

OccStress contains three evaluation families:

- Manual occupancy corruptions: `dropout`, `hole`, `misalignment`, `semantic`, and `traffic`.
- Upstream camera-only occupancy: STCOcc-style outputs, including clean and nuScenes-C-style camera corruptions.
- Upstream pointcloud-fusion occupancy: SDGOcc-style outputs, including clean and LiDAR/pointcloud stressors.

The temporal protocol names follow the paper terminology:

- `current_only`: corruption is applied to the current input state.
- `recent_burst`: corruption is applied to a recent burst of history states.
- `history_only`: corruption is applied to history states while the current state remains clean.

The default forecasting window is `H4_F6`: four historical occupancy states and one current occupancy state are used to predict six future occupancy states.

## File Format

Occupancy payloads are stored in tar shards. Each shard member preserves its original relative path, for example:

```text
occ/manual/dropout/easy/<scene>/<sample_token>/labels.npz
occ/upstream/camera_only/stcocc/clean/<scene>/<sample_token>/labels.npz
occ/upstream/pointcloud_fusion/sdgocc/clean/<scene>/<sample_token>/labels.npz
cache/manual/misalignment/<...>
```

Use `manifests/shards.jsonl` for shard-level metadata and `manifests/shard_members/*.lst` for exact tar member lists. Use `manifests/source_files.jsonl` and `manifests/source_checksums.sha256` for the original source payload manifest.

Event metadata is grouped into JSONL files under `events/`. Each line has the form:

```json
{"path":"events/.../sample.json","data":{...}}
```

Protocol files under `protocols/` are evaluation index files consumed by the benchmark evaluation code.

## Integrity

The release includes:

- `manifests/files.jsonl`: file list and SHA-256 hashes for the HF staging layout.
- `manifests/checksums.sha256`: SHA-256 checksums for the HF staging layout.
- `manifests/source_files.jsonl`: source payload file list before sharding.
- `manifests/source_checksums.sha256`: source payload SHA-256 checksums.
- `manifests/staging_audit.json`: packaging audit.
- `manifests/source_audit_*.json`: source data audit.

The current staging audit status is `PASS`. The audit confirms that raw `occ/` and raw `cache/` directories are not duplicated in the HF staging layout, symlink count is zero, backup-name hits are zero, and old protocol names are absent from manifests.

## Responsible AI Notes

OccStress is a diagnostic evaluation dataset for autonomous-driving occupancy prediction. It does not add personally identifying labels, demographic labels, or human-subject annotations. It inherits the privacy and license constraints of the upstream nuScenes ecosystem and should be used only by users who have the necessary rights to use the underlying data.

Intended use:

- Evaluate robustness and temporal sensitivity of occupancy prediction and forecasting models.
- Compare stress-condition degradation against clean validation protocols.
- Support reproducible analysis of occupancy-state corruption and upstream perception failure modes.

Out-of-scope use:

- Training or deploying safety-critical driving systems directly from these stress annotations.
- Inferring personal, demographic, or behavioral attributes.
- Treating synthetic stress conditions as a complete model of real-world safety risk.

Known limitations:

- This is an evaluation-only validation benchmark; it does not include a training split.
- It covers selected occupancy-state and upstream-perception stressors, not all possible autonomous-driving hazards.
- Metrics should be interpreted as robustness diagnostics, not as standalone safety guarantees.
- The included `sample/OccStress_mini.tar` is a compact reviewer subset; full benchmark results should use the complete shard/protocol release. The sampling procedure is documented in `sample/README.md`.

## Licensing

This release is provided for non-commercial research evaluation, subject to the upstream nuScenes and occupancy-annotation terms. Users must obtain and comply with any required upstream dataset licenses before using the benchmark. See `LICENSE` for the release terms and restrictions.

## Citation

The anonymous citation placeholder is in `CITATION.cff`. Replace it with the final accepted paper citation before public release.