Holi-Spatial

Evolving Video Streams into Holistic 3D Spatial Intelligence

Paper Code Data (coming soon) Benchmark (coming soon)

13K+

Scenes

1.3M+

QA Pairs

300K+

3D BBoxes

Yuanyuan Gao^{1, 2*}, Hao Li^{2, 5*}, Yifei Liu^{1, 6*}, Xinhao Ji^{1, 4*}, Yuning Gong^{1, 7*}, Yuanjun Liao⁷, Fangfu Liu⁸, Manyuan Zhang⁹, Yuchen Yang¹⁰, Dan Xu¹¹, Xue Yang³, Huaxi Huang¹, Hongjie Zhang¹, Ziwei Liu⁵, Xiao Sun¹, Dingwen Zhang^2,†, Zhihang Zhong^3,†

* Equal Contribution · † Corresponding Authors

¹ Shanghai AI Lab ² Northwestern Polytechnical University ³ Shanghai Jiao Tong University ⁴ Peking University ⁵ Nanyang Technological University ⁶ Beihang University ⁷ Sichuan University ⁸ Tsinghua University ⁹ The Chinese University of Hong Kong ¹⁰ Fudan University ¹¹ Hong Kong University of Science and Technology

c9ab

Clip 1

0645-1

Clip 2

07b-1

Clip 3

0572-1

Clip 4

0570-1

Clip 5

0602-1

Clip 6

394a5-1

Clip 7

0567-1

Clip 8

Unified Video Stream Modeling

Integrates continuous egocentric clips into coherent spatial trajectories for holistic scene understanding.

Grounding + Depth Coupling

Jointly analyzes object grounding and depth estimation to expose cross-view consistency and failure cases.

Benchmark-Driven Evaluation

Connects visual evidence to downstream spatial QA, enabling robust and interpretable model comparison.

Detail Showcase

Geometric Visualization

Visualize scene geometry through synchronized 3D grounding and depth estimation. Swipe or click dots to explore different scenes.

Swipe or click dots to switch · Videos play in sync

3D Visualization

Mesh Comparison

Compare lightweight base-plus-delta scene geometry across methods. Drag to rotate, scroll to zoom.

Spatial QA

Spatial QA Demonstration

Comprehensive spatial intelligence QA data covering camera motion, object relations, and camera-object interactions — capturing diverse spatial reasoning tasks across real indoor scenes.