MoSE3 recovers per-pixel world-space SE(3) via 3D point tracks and differentiable Horn fit
zhenjun_zhao · x · 2026-10-08
MoSE3 learns world-space SE(3) at every pixel: it adds a tracking branch on top of π3 to produce 3D point tracks and rigidity embeddings, then recovers per-pixel SE(3) motion through a differentiable Horn fit. Authors include Yilun Du and Qianqian Wang.
Related event: MoSE3: Pixel-Level SE(3) Estimation from Monocular Video(2 posts)→
More from Research
- Chrome silently ships a 425KB neural echo cancellation model; reverse-engineer finds 5.4dB gain — chadwallacehart · 2026-10-08
- Snorkel scales open benchmark grants 10x to $30M, adds red team for benchmarks — ajratner · 2026-10-08
- AI Theorem Prover for Hardware Formal Verification Presented at REBASE 2026 Industry Track — satnam6502 · 2026-10-08
- 1985 practitioner sides with Schmidhuber: backprop history isn't the US narrative — SchmidhuberAI · 2026-10-08
- NVIDIA's PivotOPD teaches agents to prevent and recover from pivotal early mistakes — NVIDIAAI · 2026-10-08
- Scott Aaronson's 'The Mathocalypse': math education and research in the age of capable AI — TMWNN · 2026-10-08