AevaScenes Python SDK

A Dataset and Benchmark for FMCW LiDAR Perception

<img src="https://img.shields.io/badge/aevascenes-v0.1-green?style=flat-square" alt="AevaScenes Versions">
<img src="https://img.shields.io/badge/python->=3.10-green?style=flat-square" alt="Python Versions">
<img src="https://img.shields.io/github/license/aevainc/aevascenes?style=flat-square&color=orange" alt="License">
<img src="https://img.shields.io/badge/license-AevaScenes%20Dataset%20License-orange?style=flat-square" alt="License">







<a href="https://scenes.aeva.com/"><strong>Website</strong></a> |
<a href="https://scenes.aeva.com/dataset"><strong>Dataset</strong></a> |
<a href="https://scenes.aeva.com/download"><strong>Download</strong></a> |
<a href="#license"><strong>License</strong></a> |
<a href="#citation"><strong>Citation</strong></a>

AevaScenes is a comprehensive multi-modal dataset designed to advance research in FMCW (Frequency-Modulated Continuous Wave) LiDAR perception. The dataset features synchronized data from 6 FMCW LiDARs and 6 high-resolution cameras mounted on a vehicle, captured across diverse urban and highway environments in the San Francisco Bay Area.

Instantaneous Velocity Measurements

AevaScenes features the longest-range FMCW LiDAR data ever released to the public, delivering over 400 meters of range with per-point velocity—enabling researchers and developers to explore perception capabilities beyond the limits of traditional datasets.

Ultra-Long Range Detections

For the first time, FMCW LiDAR enables detections up to 400 meters by measuring instantaneous radial velocity per point directly at the sensor. Aeva's Doppler-based approach enhances long-range perception, enabling earlier detection of moving objects and improving overall safety for autonomous navigation.

High Fidelity Perception Labels

AevaScenes provides annotations—including 3D bounding boxes, lane lines, and semantic segmentation—at ranges up to 400 meters. This unprecedented depth of annotation empowers research in long-range perception, planning, and tracking beyond the limits of existing datasets.

High Dynamic Range Reflectivity

Unlike traditional LiDARs, FMCW LiDAR delivers high dynamic range with negligible blooming around retroreflective surfaces such as road signs, botts-dots and license plates, resulting in sharper object boundaries and more accurate perception, even in challenging high-reflectivity scenarios.

Interactive Web Visualizer

A small subset of the highway/city/day/night sequences are available to see using the web visualizer here AevaScenes Web Visualizer.

Download Dataset

1. Visit scenes.aeva.com/download.
2. Register and agree to the license terms.
3. We will email you the signed_urls.txt file which contains the download links.
4. Download the dataset using the provided download script and signed_urls.txt:

# Download the dataset (after signing up and obtaining access)
mkdir -p data/aevascenes_v2
bash scripts/download_dataset.sh --url-file signed_urls_train.txt --output data/aevascenes_v2
bash scripts/download_dataset.sh --url-file signed_urls_validation.txt --output data/aevascenes_v2
bash scripts/download_dataset.sh --url-file signed_urls_test.txt --output data/aevascenes_v2

# Extract the dataset
cd data/aevascenes_v2/   # split = train | validation | test
for split in train validation test; do [ -d "$split" ] && for f in "$split"/*.tar.gz; do tar -xzf "$f" -C "$split"; done; done

Getting Started

Please see Dataset.md for details about the dataset and schema.

Please see Getting Started.md to get started with using the Python SDK. Here's a quick overview.

# List available sequences
python examples/visualize_aevascenes.py --dataroot data/aevascenes_v2 --list-sequences

# Visualize both raw and compensated point clouds
python examples/visualize_aevascenes.py --dataroot data/aevascenes_v2 --sequence-uuid <UUID> --pcd-type raw_and_compensated

# Visualize a single sequence (defaults: compensated clouds, vehicle frame, velocity coloring)
python examples/visualize_aevascenes.py --dataroot data/aevascenes_v2 --sequence-uuid <UUID>

# Project LiDAR points onto camera images
python examples/visualize_aevascenes.py --dataroot data/aevascenes_v2 --sequence-uuid <UUID> --color-mode semantic --project-points

# Raw point clouds in world frame, LiDAR-only
python examples/visualize_aevascenes.py --dataroot data/aevascenes_v2 --sequence-uuid <UUID> --pcd-type raw --coordinate-frame world --no-images

Options: --color-mode (velocity | reflectivity | semantic), --pcd-type (compensated | raw | raw_and_compensated), --coordinate-frame (vehicle | world), --project-points, --no-images, --image-downsample-factor (1 | 2 | 4 | 8), --no-keep-alive, --list-sequences.

License

The AevaScenes dataset is provided under the AevaScenes Dataset License Agreement for non-commercial use only. The AevaScenes Python SDK is licensed under the MIT License.

• Dataset License: AevaScenes Dataset License
• Code License: MIT License

Citation

If you use AevaScenes in your research, please cite our work using the following BibTeX.

@misc{aevascenes,
  title        = {AevaScenes: A Dataset and Benchmark for FMCW LiDAR Perception},
  author       = {Narasimhan, Gautham Narayan and Vhavle, Heethesh and Vishvanatha, Kumar Bhargav and Reuther, James},
  year         = {2025},
  url          = {https://scenes.aeva.com/},
}

Contributing

We welcome contributions to improve the AevaScenes dataset and toolkit! Please see our contributing guidelines and submit pull requests for:

• Requests for data diversity.
• Bug fixes and performance improvements.
• New visualization features.
• Dataset utilities and analysis tools.

Support

• Website: scenes.aeva.com
• Issues: GitHub Issues
• Email: research@aeva.ai