Online real-estate listings are a rich but underused source of indoor imagery: they show many rooms, but from few viewpoints and often with texture-poor areas. The authors of Holi-Spatial 360 propose a generative reconstruction framework that turns such sparse panoramas into fully annotated, multi-room 3D environments. Their method first converts feed-forward reconstruction priors into reliable cross-view constraints via correspondence-guided geometric optimization, then builds a coherent 3D Gaussian Splatting representation that acts as spatial memory. Adapted single-image and video priors restore appearance and fill in continuous trajectories while preserving observed content.
The reconstructed scenes support arbitrary RGB-D rendering, which feeds an automated curation pipeline combining end-to-end 3D grounding with vision-language-model (VLM) verification. This pipeline produced Holi360, a dataset with over 1,000 multi-room 3DGS scenes, more than 100,000 instance annotations, and 2 million spatial question-answer pairs. The authors also introduce the Holi360 benchmark for evaluating VLM spatial understanding in real-world, multi-room indoor environments.
Experiments reported in the paper show that Holi-Spatial 360 achieves superior reconstruction and curation quality under sparse panoramic inputs compared to existing approaches. Fine-tuning VLMs on Holi360 substantially improves their spatial reasoning capabilities, suggesting that synthetic 3D scenes derived from ordinary listing photos can serve as effective training data for spatial intelligence. The work is available on arXiv under the identifier 2610.19502.