R3D2: Realistic 3D Asset Insertion via Diffusion for Autonomous Driving Simulation
Paper • 2506.07826 • Published
Datasets:
Request research access to R3D3
Requests are reviewed individually. Register with Waymo and accept its dataset terms before applying. Describe your research and intended use.
Confirm your existing Waymo registration and acceptance at https://waymo.com/open/terms/. Submitting this form does not register you with Waymo. Your responses are used to assess your access request.
Log in or Sign Up to review the conditions and access this dataset content.
R3D3
Paired images for R3D2: 29,246 training pairs and 9,091 validation pairs, approximately 74 GB. FRONT camera only, with blacklisted sequences excluded. Original 1920×1280 JPEGs are preserved unchanged.
Non-commercial use only. R3D3 was made using the Waymo Open Dataset, provided by Waymo LLC under the Waymo Dataset License Agreement for Non-Commercial Use, and your access and use of this work are governed by its terms and conditions. Redistribute only to recipients registered with Waymo who have accepted its terms, subject to the agreement's other conditions. The complete agreement is in LICENSE.txt.
Access
Access is restricted and manually approved case by case to support reproduction, direct comparisons with R3D2, and related non-commercial research. Approval is not guaranteed. Please provide your details and describe your intended use in the request form. Each researcher needing access should submit their own request; do not share access tokens or credentials.
You must already have registered with Waymo and accepted its dataset terms. All use and further sharing remain subject to the Waymo agreement.
Download
After approval, authenticate and download the dataset folder:
pip install huggingface_hub
hf auth login
hf download bertaveira/R3D3 --repo-type dataset --local-dir R3D3
cd R3D3
sha256sum -c SHA256SUMS
This downloads the original JPEG pairs and metadata directly, approximately 74 GB. No archive extraction is required. Then set the local path as shown below.
Format
metadata.json contains root, train, and val. Each sample has:
wo_shadow: input image path, relative toroot.gt: target image path, relative toroot.seq: sequence ID.crops: actor boxes[center_x, center_y, width, height]in original image pixels.
Images are under images/<sequence>/input/ and images/<sequence>/target/.
For actor metrics after the 1920×1080 center crop, subtract 100 from each box's
center y and clip to image bounds. The training loader crops and normalizes the
images automatically.
Use
After extracting, run this once inside the dataset folder to set its local path:
python - <<'PYTHON'
import json
from pathlib import Path
p = Path("metadata.json")
metadata = json.loads(p.read_text())
metadata["root"] = str(Path.cwd())
p.write_text(json.dumps(metadata, indent=2) + "\n")
PYTHON
Then, inside the R3D2 code repository, after uv sync:
uv run r3d2/scripts/train.py R3D2 --data.data-path /path/to/R3D3
To check the transfer, run sha256sum -c SHA256SUMS inside the dataset folder.
The path-configurable metadata.json is excluded from the checksum list.
- Downloads last month
- 55