Reproduce “S4VY: Segment Anything in Feed-Forward 4D Visual Geometry” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jingdong Zhang, Xin Li, Jan Kautz, Wenping Wang, Chris Choy

S4VY segments persistent instances across sets of RGB observations, including sparse, shuffled views. Its visual-geometric backbone contextualizes observations, while a Space-Time Query Decoder assigns each object a persistent query rather than linking frame masks afterward. On LVOS Hard, it scores 0.588 J&F with 16 appearance-diverse shuffled observations, versus 0.462 for GLEE and 0.248 for IGGT4D; this benchmark result does not establish order-independence.

Key insights

  • Shared geometric and cross-observation features can support object identity without relying on neighboring-frame propagation or chronological order.
  • The system unifies exhaustive segmentation and prompted selection while keeping prompt-conditioned predictions from changing prompt-free instance discovery.
  • Adaptive search over observation subsets can reduce language-grounding effort compared with applying fixed sliding windows across every frame.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.