View-of-Delft (VoD)
The View-of-Delft (VoD) dataset is a novel automotive dataset recorded in Delft, the Netherlands. It contains 8600+ frames of synchronized and calibrated 64-layer LiDAR-, (stereo) camera-, and 3+1D (range, azimuth, elevation, + Doppler) radar-data acquired in complex, urban traffic. It consists of 123100+ 3D bounding box annotations of both moving and static objects, including 26500+ pedestrian, 10800 cyclist and 26900+ car labels. It additionally contains semantic map annotations and accurate ego-vehicle localization data.
Benchmarks for detection and prediction tasks are released for the dataset. See the sections below for details on these benchmarks.
- Detection:
An object detection benchmark is available for researchers to develop and evaluate their models on the VoD dataset. At the time of publication, this benchmark was the largest automotive multi-class object detection dataset containing 3+1D radar data, and the only dataset containing high-end (64-layer) LiDAR and (any kind of) radar data at the same time.
- Prediction:
A trajectory prediction benchmark is publicly available to enable research on urban multi-class trajectory prediction. This benchmark contains challenging prediction cases in the historic city center of Delft with a high proportion of Vulnerable Road Users (VRUs), such as pedestrians and cyclists. Semantic map annotations for road elements such as lanes, sidewalks, and crosswalks are provided as context for prediction models.
1. Install VoD Prediction Toolkit
We will use the VoD Prediction toolkit to convert the data.
First of all, we have to install the vod-devkit.
# install from github (Recommend)
git clone git@github.com:tudelft-iv/view-of-delft-prediction-devkit.git
cd vod-devkit
pip install -e .
# or install from PyPI
pip install vod-devkit
By installing from github, you can access examples and source code the toolkit. The examples are useful to verify whether the installation and dataset setup is correct or not.
2. Download VoD Data
The official instruction is available at https://intelligent-vehicles.org/datasets/view-of-delft/. Here we provide a simplified installation procedure.
First of all, please fill in the access form on vod website: https://intelligent-vehicles.org/datasets/view-of-delft/.
The maintainers will send the data link to your email. Download and unzip the file named view_of_delft_prediction_PUBLIC.zip.
Secondly, all files should be organized to the following structure:
/vod/data/path/
├── maps/
| └──expansion/
├── v1.0-trainval/
| ├──attribute.json
| ├──calibrated_sensor.json
| ├──map.json
| ├──log.json
| ├──ego_pose.json
| └──...
└── v1.0-test/
Note: The sensor data is currently not available in the Prediction dataset, but will be released in the near future.
The /vod/data/path should be /data/sets/vod by default according to the official instructions,
allowing the vod-devkit to find it.
But you can still place it to any other places and:
build a soft link connect your data folder and
/data/sets/vodor specify the
datarootwhen calling vod APIs and our convertors.
After this step, the examples in vod-devkit is supposed to work well.
Please try view-of-delft-prediction-devkit/tutorials/vod_tutorial.ipynb and see if the demo can successfully run.
3. Build VoD Database
After setup the raw data, convertors in ScenarioNet can read the raw data, convert scenario format and build the database.
Here we take converting raw data in v1.0-trainval as an example:
python -m scenarionet.convert_vod -d /path/to/your/database --split v1.0-trainval --dataroot /vod/data/path
The split is to determine which split to convert. dataroot is set to /data/sets/vod by default,
but you need to specify it if your data is stored in any other directory.
Now all converted scenarios will be placed at /path/to/your/database and are ready to be used in your work.
Known Issues: VoD
N/A