Argoverse

Argoverse
Name	Argoverse
Developer	Argo AI
Released	2019
Latest release	2021
Programming language	Python, C++
Platform	Linux, macOS, Windows
License	Proprietary / Dataset terms

Argoverse is a collection of datasets, tools, and benchmarks developed to advance research in autonomous vehicle perception, prediction, and motion planning. It provides sensor data, 3D maps, and evaluation protocols used by academic groups and industry teams to train and compare models for tasks such as object detection, trajectory forecasting, and tracking. The project intersects with efforts by research labs, robotics companies, and standards bodies to create reproducible experimentation for urban driving scenarios.

Overview

Argoverse was created by Argo AI to supply high-fidelity urban driving data compatible with work from Waymo, Cruise, Uber ATG, Bolt, and academic initiatives at Stanford University, Massachusetts Institute of Technology, Carnegie Mellon University, University of California, Berkeley, and University of Oxford. The package includes synchronized LiDAR, camera, and localization streams similar to datasets from KITTI, nuScenes, Oxford RobotCar Dataset, ApolloScape, and Cityscapes. It addresses mapping challenges akin to efforts by HERE Technologies, TomTom, and OpenStreetMap contributors, and complements simulator platforms such as CARLA, LGSVL Simulator, and AirSim.

Datasets

Argoverse provides multiple labeled collections: a stereo camera and LiDAR corpus for 3D detection, a motion forecasting dataset with multi-agent trajectories, and a high-definition (HD) map dataset for lane geometry and traffic flow. The 3D detection suite resembles work in KITTI Vision Benchmark Suite and shares annotation conventions used in datasets released by Facebook AI Research, Google Research, Microsoft Research, and Amazon Robotics. The motion forecasting portion is comparable to datasets produced by ETH Zurich, Max Planck Institute, Uber Advanced Technologies Group, Waymo Open Dataset, and NVIDIA Research. HD maps in Argoverse are structured like mapping outputs from HERE Technologies, TomTom International, Trimble, Hexagon AB, and mapping efforts by Mapbox.

Annotations include bounding boxes, object IDs, lane centerlines, lane widths, and traffic light locations reminiscent of labels in the Berkeley DeepDrive dataset, Apollo Auto releases, and datasets from Purdue University research groups. The dataset splits and evaluation protocols parallel those used by Imagenet Large Scale Visual Recognition Challenge, COCO, and PASCAL VOC communities to encourage standardized comparison.

Applications and Benchmarks

Researchers apply Argoverse data to benchmark 3D object detection, multi-object tracking, semantic segmentation, and trajectory prediction. Benchmark tasks draw comparisons to algorithms and models from PointPillars, MV3D, VoxNet, SECOND, PV-RCNN, and forecasting approaches like Social LSTM, Trajectron++, VectorNet, and DESIRE. Multi-agent prediction evaluations reference work published at conferences including CVPR, ICCV, ECCV, NeurIPS, ICLR, and RSS. The dataset supports evaluation metrics used by papers in IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Robotics and Automation Letters, and proceedings of IEEE ICRA and IEEE IROS.

Industrial applications have been demonstrated in testbeds developed by Ford Motor Company, General Motors, Toyota Research Institute, Volvo Cars, BMW Group Research, and startups incubated by Techstars. Academic projects at Princeton University, Columbia University, University of Michigan, and Cornell University have used Argoverse for coursework and thesis projects.

Tools and Software

Argoverse ships with software for visualization, data loading, and map querying implemented in Python and C++, interoperable with libraries such as PyTorch, TensorFlow, Open3D, PCL (Point Cloud Library), and NumPy. Community integrations include wrappers for ROS (Robot Operating System), plugins for RViz, and converters used with Autoware. Tooling supports use with cloud platforms from Amazon Web Services, Google Cloud Platform, and Microsoft Azure for large-scale training and with container ecosystems like Docker and orchestration via Kubernetes. Third-party projects link Argoverse to visualization tools such as Kepler.gl, Potree, and web frameworks from Node.js ecosystems.

History and Development

Argoverse was announced in 2019 following internal data collection and annotation efforts by Argo AI engineers and data scientists collaborating with partners in the autonomous driving community. The initiative followed precedents set by datasets from KITTI authors at Karlsruhe Institute of Technology and researchers at Toyota Research Institute USA. Subsequent dataset releases and software updates incorporated feedback from teams participating in workshops at NeurIPS Datasets and Benchmarks Track, CVPR Workshops, and ICRA Tutorials. Development involved annotation pipelines influenced by methods used at Labelbox, Scale AI, and open labeling projects at Carnegie Mellon University.

Impact and Adoption

Argoverse influenced benchmarks and competitions hosted by academic conferences and industry consortia, informing standards discussed at forums like IEEE Standards Association, SAE International technical committees, and policy dialogues at National Highway Traffic Safety Administration and U.S. Department of Transportation advisory panels. Its datasets have been cited in research from institutions including MIT CSAIL, Harvard John A. Paulson School of Engineering and Applied Sciences, University of Washington, Rutgers University, and companies such as Uber Technologies, Lyft, and Zoox. Argoverse fostered accelerated research in perception and prediction alongside community efforts exemplified by OpenAI collaborations and reproducibility initiatives promoted by Allen Institute for AI.

Category:Datasets