SpatialBlock: Enhancing Spatial Intelligence in LVLMs via Synthetic Block-Stacking Problem
A 15,000-example synthetic block-stacking set is used to train LVLMs on 3D spatial reasoning from 2D views.
The paper argues that real-scene spatial QA data is expensive and noisy because it needs dense geometric labels. SpatialBlock-15k instead covers projection, viewpoint changes, and structural combination in controlled block tasks. The authors say models trained on it outperform baselines and carry gains over to real-world spatial tasks. HF Daily Papers' note
The paper argues that real-scene spatial QA data is expensive and noisy because it needs dense geometric labels. SpatialBlock-15k instead covers projection, viewpoint changes, and structural combination in controlled block tasks. The authors say models trained on it outperform baselines and carry gains over to real-world spatial tasks. HF Daily Papers' note
score 4