Search papers, labs, and topics across Lattice.
SpatialBench is introduced as a comprehensive benchmark to evaluate the generalization capabilities of spatial foundation models across diverse tasks, viewpoints, domains, input densities, and hardware constraints. The benchmark comprises 19 datasets and 546 scenes across 5 spatial domains, evaluating 41 models across 6 paradigms. Results show current models struggle to generalize, highlighting the importance of full-context attention for accuracy and domain alignment/data quality for embodied/egocentric tasks; the authors also contribute a new dataset, DA-Next-5M, and baseline model, DA-Next, to address identified data gaps.
Spatial foundation models aren't as "all-round" as we thought: SpatialBench reveals surprising generalization gaps and the critical importance of domain alignment over naive data scaling.
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? Answering this overarching question requires a holistic assessment, yet current models are mainly evaluated on specific domains for which they were specifically designed or trained. Such evaluations are intrinsically limited by narrow paradigm coverage, limited scene domains, and arbitrary frame sampling, making it fundamentally difficult to assess their true generalization capabilities. To address this gap, we present SpatialBench, a cross-paradigm, domain-diverse benchmark for spatial foundation models with deterministic sampling. SpatialBench features unprecedented scale and rigorous deterministic design, comprising 19 datasets and 546 scenes across 5 diverse spatial domains. It comprehensively evaluates 41 models across 6 paradigms on 5 task suites under 4 different input density settings. Our extensive evaluation reveals that current models are not yet all-round players, and uncovers crucial insights for future advancement. Specifically, we demonstrate that full-context attention maximizes accuracy while bounded-memory strategies unlock long-sequence scalability. Moreover, our empirical evaluations in challenging embodied and egocentric tasks demonstrate that strict domain alignment and high data quality are far more critical to performance than simple dataset scaling. Furthermore, to address the largest data gap identified in our analysis, we go beyond evaluation by introducing a large-scale dataset, DA-Next-5M, and a strong baseline model, DA-Next, pushing the boundaries of spatial representation learning.
