README.md

---
license: cc-by-4.0
pretty_name: DRC CARLA Multi-Vehicle Camouflage Dataset
annotations_creators:
- machine-generated
source_datasets:
- original
task_categories:
- object-detection
size_categories:
- 1K<n<10K
tags:
- autonomous-driving
- carla
- adversarial-robustness
- differentiable-rendering
- synthetic-data
- 3d
---

# DRC CARLA Multi-Vehicle Camouflage Dataset

This repository releases the audited synthetic data and geometry assets used to
study whether one differentiable vehicle-camouflage generator transfers across
vehicle shapes. The release preserves the original split manifests, collection
protocols, audit records, and SHA-256 checksums. It is intended for reproducible
adversarial-robustness research, including the analysis of negative results.

## Release contents

| Component | Vehicles and role | Samples | Status |
| --- | --- | ---: | --- |
| Stage 2 Town10 | Audi E-Tron, Tesla Model 3, Ford Mustang, and Jeep Wrangler Rubicon (train); Toyota Prius (validation); Lincoln MKZ 2017 (development holdout) | 864 | Audited |
| Stage 3 Town10 | Nissan Patrol 2021 (outcome-sealed confirmation) | 144 | Audited; source-detector confirmation failed |
| Geometry assets | Attack-UV OBJ, binary attack mask, object-position map, and object-normal map | 7 vehicles | Hash-audited |
| Protocol records | Split manifests, collection plans, preregistrations, audits, and gate decisions | N/A | Versioned JSON |

The Stage 2 images are `1152 x 1152` and contain 144 distinct camera positions
per vehicle. The Stage 3 set contains the same 144-view grid for Nissan Patrol
2021 under Town10HD_Opt/ClearNoon. Thirty-two clean views used earlier for Nissan
mesh alignment are excluded from the released confirmation evaluation.

## Raw snapshot layout

| Hub path | Files | Exact bytes | Directory-manifest SHA-256 |
| --- | ---: | ---: | --- |
| `raw/stage2_town10_v1/` | 872 | 21,787,037,755 | `4ac82a74b0c9c8b0918b93b911997b12f6401f3cbd4b0c18914125e715faea67` |
| `raw/stage3_nissan_patrol_2021_sealed_town10_v1/` | 148 | 3,631,182,409 | `89844aa31bd9ca1803b29bb2bddc5379dc8e3ad049c6e4448ce81fbc24dc5c3f` |

Each directory has its own `SHA256SUMS`. The Stage 2 manifest covers 871
payload files (864 samples and seven JSON records); the Stage 3 manifest covers
147 payload files (144 samples and three JSON records). The checksum file itself
is excluded from the corresponding manifest and is anchored by the table above.

The immutable raw-data release is tagged `raw-v1-20260803`. For example, download
only the Stage 2 snapshot with the repository utility:

```bash
python tools/huggingface_assets.py download /path/to/release \
  --kind dataset \
  --revision raw-v1-20260803 \
  --allow 'raw/stage2_town10_v1/**'
```

Record the full commit resolved by the tag in experiment provenance rather than
relying on the moving `main` branch.

## Sample format

Each sample is a NumPy object array with seven ordered fields:

1. `image`
2. `background`
3. `mask`
4. `discrepancy`
5. `labels`
6. `camera_position`
7. `metadata`

Loading requires `numpy.load(path, allow_pickle=True)` because the record contains
heterogeneous arrays and a metadata dictionary. Only load files from a trusted
release revision and verify `SHA256SUMS` first.

## Split semantics

- Toyota is the validation vehicle used for checkpoint selection.
- Lincoln is training-unseen but has been queried during method development, so it
  is a development holdout rather than a sealed final test.
- Nissan is the frozen negative confirmation vehicle and must not be used to tune
  architectures, objectives, checkpoints, views, or seeds.
- A future positive method requires a different, newly reserved confirmation
  vehicle.

## Integrity

The accepted Stage 2 sample aggregate SHA-256 is
`a1e355159dc4425159d47bc3a88941e60877ca22bd6e8071942873725fd3c04c`.
Its split manifest SHA-256 is
`aa358e04bca98ed64dc18a72c41fa9bbcc1af0b1a9cd4483b816dae162275090`.
The Stage 3 Nissan sample aggregate SHA-256 is
`81fa63944b61bc44e28c412331d42507d2db774319db9e1372bd0772f56be3aa`;
its sealed manifest SHA-256 is
`05f30d781d083fc49ea1bbe7dfe861204f28911cecd31ba747e980e07e5db748`.

These aggregate values use the project audit procedure recorded in the included
provenance files. The two directory-level `SHA256SUMS` files additionally verify
every payload file in the raw snapshots.

## Limitations and responsible use

This is synthetic CARLA data from a limited map, weather, detector, and camera
protocol. It does not establish physical-world effectiveness or safety. The
released checkpoint did not produce a material attack effect on the sealed Nissan
vehicle, and no held-out detector evaluation was run after that gate failed.
Do not use this dataset to claim broad vehicle, detector, weather, map, or physical
transfer without new evidence.

The release is for defensive research, benchmarking, and reproducibility. Users
are responsible for complying with applicable laws and deployment policies.

## License and attribution

The dataset and project-generated annotations are released under CC BY 4.0.
The scenes and vehicle assets are derived from CARLA 0.9.15. CARLA-specific assets
are distributed by the CARLA project under CC BY; this release modifies them by
exporting geometry, constructing attack UVs, rasterizing masks, and computing
object-space position/normal maps. Cite and attribute the CARLA project when using
these materials: https://github.com/carla-simulator/carla

The license does not cover unrelated third-party detector or foundation-model
weights, which are intentionally excluded.