V2X-Radar

What It Is

V2X-Radar is a real-world cooperative perception dataset with 4D radar.
It was accepted as a NeurIPS 2025 Spotlight according to the official repository and arXiv page.
The dataset covers vehicle-side and infrastructure-side sensing.
It adds 4D radar to the cooperative perception setting where earlier datasets focused mostly on camera and LiDAR.
It supports cooperative, roadside-only, and vehicle-only perception benchmarks.
It is a dataset and benchmark codebase, not a single detector architecture.

Collect synchronized data from a connected vehicle platform and an intelligent roadside unit.
Equip both sides with 4D radar, LiDAR, and multi-view cameras.
Provide real-world scenes where cooperative perception can extend range and reduce occlusion.
Include adverse lighting and weather so radar value can be measured in cooperative settings.
Split the dataset into task-oriented subsets for V2X, infrastructure, and vehicle perception.
Provide codebase support for several cooperative perception models.

Dataset input: 4D radar point clouds.
Dataset input: LiDAR point clouds.
Dataset input: multi-view camera images.
Dataset input: calibration and metadata for vehicle and roadside agents.
Labels: 3D bounding boxes across five categories.
Benchmark output: cooperative or single-agent 3D object detection predictions and AP metrics.

The dataset contains 20K LiDAR frames.
It contains 40K camera images.
It contains 20K 4D radar frames.
It provides about 350K annotated 3D boxes.
Subsets include V2X-Radar-C for cooperative perception, V2X-Radar-I for roadside perception, and V2X-Radar-V for vehicle perception.
The official codebase supports cooperative models such as F-Cooper, CoAlign, V2X-ViT, CoBEVT, and HEAL.

The arXiv page reports comprehensive benchmarks across the three sub-datasets.
Scenarios include sunny and rainy weather plus daytime, dusk, and nighttime.
The repository includes data in OPV2V and KITTI-style formats.
Cooperative AP is reported by distance ranges such as 0-30 m, 30-50 m, and 50-100 m.
Evaluation can compare late fusion, intermediate fusion, roadside-only, and vehicle-only settings.
Dataset revisions and pretrained weights were released in January 2026 according to the repository changelog.

First large real-world cooperative perception dataset centered on 4D radar.
Directly useful for studying infrastructure radar under rain and night conditions.
Includes both single-agent and cooperative benchmarks.
Public codebase lowers the cost of reproducing V2X baselines.
Radar point clouds add velocity and weather robustness to V2I perception.
Strong fit for bounded operational domains with fixed infrastructure.

Roadside geometry differs from airport aprons, stands, gates, and service roads.
Cooperative perception depends on calibration, time synchronization, communication latency, and packet loss.
Dataset boxes do not cover airport-specific objects or aircraft clearance envelopes.
4D radar point clouds may lose information compared with raw radar tensors.
Benchmarks may reward average AP while hiding rare safety-critical classes.
Infrastructure sensors introduce cybersecurity and maintenance requirements.

Very high relevance because airports can deploy fixed roadside sensors and private 5G.
Infrastructure 4D radar is attractive for rain, night, fog, de-icing, and occlusion around aircraft.
V2X-Radar provides the closest public analogue for vehicle-plus-infrastructure 4D radar perception.
Airport adaptation needs stand-side sensor placement studies and airside-specific labels.
Cooperative outputs should enhance, not gate, the vehicle's onboard safety autonomy.
Best use is a benchmark template for an airside V2I radar dataset.

Reuse the dataset schema ideas for airport data: vehicle node, infrastructure node, calibration, synchronized radar/LiDAR/camera.
Add latency traces and packet-loss logs to any airside collection because V2X timing is safety-relevant.
Keep single-agent baselines so infrastructure value can be quantified.
Track per-class and per-zone performance, not just aggregate AP.
Evaluate under night/rain and with aircraft occlusions specifically.
Separate safety-critical onboard perception from cooperative enhancement in architecture documents.