SpatialBench: Is Your Spatial Foundation Model an All-Round Player?

SpatialBench is a new large-scale benchmark for spatial foundation models, spanning 19 datasets, 546 scenes, 41 models, and 5 spatial domains.

Its evaluations show that many models still struggle to generalize across domains, tasks, and input densities.

The study also finds that full-context attention and bounded-memory methods can improve performance.

To support future progress, the researchers released DA-Next-5M and the DA-Next baseline model.