RefSpatial-Bench: A Benchmark for Spatial Referring Tasks in Indoor and Outdoor Scenes

RefSpatial-Expand-Bench extends evaluation for spatial referring tasks to include both indoor and outdoor environments. The associated paper, RoboRefer, was accepted to the NeurIPS 2025 conference. It was released by BAAI and last updated on October 23, 2025.

Use Cases

• Benchmarking model performance on spatial referring tasks based on the described indoor scenes like factories and stores.
• Evaluating the generalization of vision-language models based on the introduced outdoor scenarios such as streets and parking lots.
• Training or fine-tuning robotic perception systems based on the spatial referring tasks central to the benchmark.

Strengths

• Expands evaluation scope to include both indoor and outdoor scenes, as stated in the description.
• Associated with a peer-reviewed paper accepted to NeurIPS 2025, indicating academic validation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.

Provenance

Source

BAAI

Freshness

Last updated 2025-10-23 07:32:11; freshness should be verified.