S2Tok Proposes Persistent Spatial Tokens for Streaming 3D Gaussian Reconstruction
zhenjun_zhao · x · 2026-10-09
S2Tok: Streaming 3D Gaussian Reconstruction with Persistent Spatial Tokens
- Problem: Streaming 3D reconstruction needs a persistent scene state that absorbs new evidence and stays renderable; maintaining latent spatial tokens online was open.
- Method: A feed-forward framework distinguishing updates to existing representation from selective expansion — a spatially informed transformer integrates each observation into persistent scene tokens, while a learned admission module expands representation selectively to limit redundant storage. A hierarchical decoder plus Gaussian head outputs non-pixel-aligned 3D Gaussians, enabling novel-view rendering without caching past frames.
- Results: Competitive streaming rendering quality across four benchmarks with compact Gaussian representations, supporting latent spatial tokens as persistent computational state for online 3D reconstruction.
More from Research
- Are URM and Universal Transformers the forgotten architecture beating standard LLMs? — moschles · 2026-10-09
- Researcher: Use AI to Rewrite Machine-Generated Math Proofs Into Human-Readable Forms — jd_pressman · 2026-10-09
- AutoScientist's two-agent checklist loop auto-audits every training example — sarahookr · 2026-10-09
- NeurIPS GenAI4Health oral: retrieval-based medical fact-checking fails in ways bigger models can't fix — mdredze · 2026-10-09
- Bigger models, more reasoning, better sources won't fix medical fact-checking, researchers say — mdredze · 2026-10-09
- DEX best abstract: top LLMs catch many physician diagnostic errors, but big gaps remain — mdredze · 2026-10-09