G3T Project Page: Gravity-Aligned Pointmaps Beat Camera-Centric Frames Versus VGGT
rsasaki0109 · x · 2026-10-04
The G3T project page by Bharath Raj Nagoor Kani and Noah Snavely (Cornell) details their gravity-aligned coordinate frame approach.
- Core idea: instead of camera-centric pointmaps, predict upright, gravity-aligned ones that exploit strong structural cues in real-world scenes
- Gravity-aligned frames share a common vertical axis across viewpoints, reducing rotational degrees of freedom needed to relate pointmaps
- G3T outputs upright pointmaps regardless of input orientation; G3T-Long leverages this for robust long-sequence 3D reconstruction
- Interactive comparisons on three self-captured scenes show G3T's horizontal surfaces at uniform height while VGGT's show tilt gradients
arXiv paper and code are linked.
More from Research
- Scientists open PRs on HF datasets: BIDS builder for neuroimaging and biosignals — evijit · 2026-11-20
- Hugging Hub science data formats mapped: AnnData leads genomics at 17% — evijit · 2026-11-20
- Fully AI-designed molecule Dyno-hc9: 7 mutations picked by deep learning, crosses blood-brain barrier — IAmSamFin · 2026-11-20
- Fine-tuned Qwen3-VL 8B reads camera shots at 69% accuracy, beating Opus and Gemini — ravisparikh · 2026-11-20
- Victor Taelin's Bend2 talk from SF Lean Meetup explains the rationale behind the language design — pbaylies · 2026-11-20
- Protocol Labs Launches Neuro Atlas: Companies, Milestones and BCI Funding Mapped — melnykowycz · 2026-11-20