Benchmark RadarSpatialBench
SpatialBench is a deterministic, density-aware benchmark for spatial foundation models, spanning 19 datasets, 546 scenes, and five spatial domains. It evaluates 41 models across six paradigms on five task suites—depth, camera pose, trajectory, point-cloud reconstruction, and long-sequence streaming—under four input density settings with precomputed and pinned test frames.
- Released
- 2026-05-26
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Spatial foundation models are typically evaluated only on domains they were designed for, making cross-domain generalization difficult to assess. SpatialBench provides a controlled protocol with fixed sampling and multiple density settings, enabling a holistic comparison of generalization across viewpoints, scene domains, and hardware constraints.
Motivation
While spatial foundation models have demonstrated impressive performance on standard datasets, a critical question remains: are they truly all-round players capable of generalizing robustly across diverse downstream tasks, arbitrary viewpoints, shifting scene domains, varying input densities, and specific hardware constraints?
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.