📦 Instok3D: Object-Based 3D Reconstruction
The Instok3D framework shifts from representing scenes through sparse primitives to structured groups of tokens. By utilizing two-layer factorization (Instance Tokens for entities and Anchor Tokens for geometry), the model works with unposed images, allowing for the segmentation and manipulation of objects in a single pass.
🌍 The transition to an object-oriented representation simplifies high-level editing and semantic search in 3D scenes, which is critical for robotics and AR.
👤 It is now possible to create and edit complex 3D scenes simply by manipulating objects, rather than manually adjusting every polygon.
Source 1: https://yoomimi.github.io/instok3d Source 2: https://arxiv.org/abs/2606.29513