Cornell's G3T Predicts Gravity-Aligned Pointmaps for Robust Long-Sequence 3D Reconstruction
rsasaki0109 · x · 2026-10-04
Cornell University introduces G3T (Gravity Grounded Geometry Transformer) and the G3T-Long pipeline for long-sequence 3D reconstruction.
- Feed-forward methods like VGGT predict pointmaps in camera-centric frames, requiring extra rotational alignment across viewpoints
- G3T instead predicts upright, gravity-aligned pointmaps regardless of input orientation: floors, stairs and benches sit at consistent heights, and all views share a common vertical axis, cutting rotational degrees of freedom
- G3T-Long exploits this uprightness for robust long-sequence reconstruction
- The project page includes interactive comparisons showing VGGT's tilted surfaces versus G3T's level ones
Paper, code and project page are available.
More from Research
- Humans are vastly more sample-efficient than LLMs, and finding why could rival attention — burny_tech · 2026-10-04
- Tsinghua and Peking University Probe Whether LLMs Think Beyond Language in Nature MI — jiqizhixin · 2026-10-04
- Judea Pearl Announces Second Edition of The Book of Why, Due October 20 — yudapearl · 2026-10-04
- Nature study: policies that spread cooperation concentrate benefits among the well-connected — arjunrajlab · 2026-10-04
- Large-number subtraction haunts all softmax kernels; maybe fwd-bwd consistency is all we need — YouJiacheng · 2026-10-04
- Bumble bees show goal-directed tool use without training, a first in insects, Science study finds — anselm · 2026-10-04