Download
The full dataset is 1.3 TB, released under the MIT license. It is split into parts, so you only need to download the modalities you use.
Parts
Each part is split into a train and a test archive. Start with Base; the other parts extend it in place.
| Part | Contents | Train | Test |
|---|---|---|---|
| Base | Front camera, GNSS, IMU, weather, anomaly labels | 137 GB | 86 GB |
| Camera extended | Left, right and rear cameras with instance segmentation and anomaly labels | 427 GB | 267 GB |
| LiDAR | Point clouds with anomaly and semantic segmentation labels | 227 GB | 138 GB |
| KITTI | LiDAR labels in KITTI format | 22 GB | 12 GB |
| Depth | Depth maps for all cameras | 74 GB | 48 GB |
With the dev kit
The carlanomaly Python package downloads the part each dataset class needs on first use, verifies its checksum and resumes interrupted transfers.
pip install git+https://github.com/carlanomaly/devkit
from carlanomaly.datasets.rgb import RGBDataset
# Fetches the base training archive into ./data on first use
rgb = RGBDataset("./data", split="train", direction="front")
frame = rgb[0] # FloatTensor (T, 3, H, W) in [0, 1]
Manually
Download and unpack the base dataset:
cd my/data/dir/
wget https://data.carlanomaly.de/v1/carlanomaly-base-train.tar.gz
tar -xzf carlanomaly-base-train.tar.gz
Then add further parts into the same directory:
cd my/data/dir/
wget https://data.carlanomaly.de/v1/carlanomaly-<name>-<split>.tar.gz
tar -xzf carlanomaly-<name>-<split>.tar.gz