The Year in AI Papers
The year/Independent research

Paper 2605.27367

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

Haosong Peng, Hao Li, Jiaqi Chen, Yuhao Pan, Runmao Yao, Yalun Dai, Fushuo Huo, Fangzhou Hong, Zhaoxi Chen, Haozhao Wang, Dingwen Zhang, Ziwei Liu, et al.

Published
May 2026
Research lab
Independent
Citations
0
GitHub
119 stars

01 In brief

Summary

SpatialBench is a new benchmark for evaluating spatial foundation models across diverse domains, input densities, and model paradigms.

It includes 19 datasets, 546 scenes, 41 model variants, and 6 paradigms, using a deterministic multi-density sampling protocol (single, sparse, medium, dense).

Key findings reveal that full-context attention models achieve the highest accuracy but suffer from memory limits, while bounded-memory models enable long-sequence reconstruction at the cost of accuracy.

Data quality is more critical than data volume, and egocentric and wrist-view domains remain the most significant out-of-distribution failure modes.

To address this gap, the authors introduce DA-Next-5M, a dataset with 5.5M frames from 22K egocentric and wrist-view scenes, and DA-Next, a model that improves depth estimation by 47-59% and pose estimation by 3.1-5.5% over DA3-Giant on sparse/medium inputs.

The benchmark also reveals that test-time training (TTT) methods provide gains mainly on dense long sequences, and that injecting ground-truth depth priors consistently improves depth accuracy, while camera pose priors yield inconsistent benefits.

02 From the paper

Abstract

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.