RefAV: Referring Expressions and 3D Bounding Boxes for Autonomous Driving Scenarios

RefAV is a scenario mining dataset built on Argoverse 2, containing language prompts paired with corresponding 3D bounding boxes. The dataset is used for the 2026 Argoverse 2 Scenario Mining Challenge and contains between 10,000 and 100,000 records. It was uploaded by CainanD to Hugging Face on February 17, 2026.

Use Cases

• Training models for 3D object localization based on natural language referring expressions.
• Evaluating the performance of language-conditioned perception systems in driving scenarios.
• Benchmarking progress on the 2026 Argoverse 2 Scenario Mining Challenge tasks.

Strengths

• Built on the established Argoverse 2 autonomous driving dataset, suggesting a solid foundational source.
• Contains between 10,000 and 100,000 records, indicating a moderate to large scale.
• Specifically designed for a 2026 benchmark challenge, implying a defined evaluation purpose.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown within the 10K-100K range, which may limit suitability assessment.
• The license is listed as CC-BY-NC-SA-4.0, which restricts commercial use.

Provenance

Source

Argoverse 2 dataset

Collection Method

Derived through scenario mining; likely contains annotations linking language to 3D boxes.

Time Range

The underlying Argoverse 2 data likely covers specific temporal ranges, but this is not specified.

Freshness

Last updated 2026-02-17 20:37:01; freshness should be verified.

Geography

The underlying Argoverse 2 data likely covers specific geographic areas, but this is not specified.