Poincar3: 3D Self-Supervised Learning via Multi-View Self-Distillation
rsasaki0109 · x · 2026-11-07
The post introduces the paper Emergent Multi-View Geometry Through Self-Distillation, proposing Poincar3, a 3D self-supervised learning (SSL) method trained through multi-view self-distillation.
Without any supervision, Poincar3's attention maps alone can produce reliable point tracks, showing emergent multi-view geometric matching ability. The name honors Henri Poincaré's argument that a being without motion cannot understand 3D space — motivating the use of multi-view motion data (an idea originating from SeeSE(3), Chen et al., 2026).
More from Research
- Fudan's D3D diffusion drafter enables async speculative decoding, up to 7.48x speedup on Qwen3 — pmttyji · 2026-11-07
- AISTATS PC doubles down: conference exists to filter correct papers, not give feedback — qberthet · 2026-11-07
- A misplaced pen dip in 1916 gave us the crystal growth method behind every chip — aakashgupta · 2026-11-07
- Near-exact cache matches cause 2-7x more wrong replies than the verified gray zone — Reasonable_Royal_621 · 2026-11-07
- RLAF: Nace AI trains decision models with a reasoning model as judge — nischay_twt · 2026-11-07
- ICLR 2027 reviews criticized as unusually shallow: generic bullets, low scores without justification — Charming-Fail-772 · 2026-11-07