Reproduce “DynaTok: Token-Based 4D Reconstruction from Partial Point Clouds” 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
Weirong Chen, Keisuke Tateno, Hidenobu Matsuki, Michael Niemeyer, Daniel Cremers, Federico Tombari
Researchers from TU Munich and Google developed DynaTok, a point-based framework that reconstructs complete 4D scenes from sequences of partial, unordered, and correspondence-free point clouds. It leverages latent temporal aggregation and a conditional global decoder to achieve higher completeness and normal consistency on object and scene-level benchmarks compared to prior methods, even with noisy real-world data.
Key insights
- Effective 4D reconstruction from partial point clouds requires structured temporal aggregation in a latent token space to compensate for missing geometry and establish a coherent canonical representation.
- Decoupling time-invariant scene structure and time-varying deformations within a shared latent space, via a residual formulation, simplifies the modeling of complex dynamics.
- A conditional flow-matching decoder can generate dense and complete 4D point clouds by integrating noise samples through a learned velocity field, conditioned on the aggregated latent tokens.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.