First video stitching method for 5+ freely moving cameras accepted at NeurIPS 2026
kwangmoo_yi · x · 2026-11-10
Ifergane et al. from Ben-Gurion University present Video Stitching from Multiple Moving Cameras (NeurIPS 2026), the first method handling more than two freely moving cameras — with visuals that look genuinely good.
Key contributions:
- Regularization-free formulation for joint alignment across views and time;
- An annotated multi-view dataset (MMCD) and a benchmark for unconstrained camera motion;
- A variance-based geometric metric for multi-view alignment.
Existing methods mostly support only two cameras; StabStitch++ reaches N=3 but requires a fixed camera ordering and degrades when cameras swap. Per-frame image-stitching methods produce temporally incoherent results, and adding stabilization helps but isn't enough.
Related event: Video Stitching from 5 Freely Moving Cameras Accepted to NeurIPS 2026(2 posts)→
More from Multimodal
- Early access user showcases MiniMax H3-Image with surreal floating teacup dreamscapes — SimplyAnnisa · 2026-11-10
- Creator shows headless motion animation control side-by-side, clarifies it's not Fable 5.2 — RichardsonDx · 2026-11-10
- Fable 5.2 stuns with code-generated cinematic 3D journey through history of science — DeryaTR_ · 2026-11-10
- Seedance 2.5 'explains' how the pyramids were built: giants hauling bricks — FinanceYF5 · 2026-11-10
- MiniMax H3 Image early access: exploded CAD-style drone blueprints impress testers — umesh_ai · 2026-11-10
- GPT Image 2 still unmatched at depicting genuinely strange creatures — liminal_bardo · 2026-11-10