UniQueR: Unified Query-based Feedforward 3D Reconstruction

Arxiv here

Abstract:

Existing feedforward models such as DUSt3R, VGGT, and AnySplat predict per-pixel point maps or pixel-aligned Gaussians, which remain fundamentally 2.5D and limited to visible surfaces. UniQueR instead learns a compact set of 3D anchor points that act as explicit geometric queries in global 3D space, each spawning a set of 3D Gaussians for differentiable rendering. On Mip-NeRF 360 and VR-NeRF it surpasses state-of-the-art feedforward methods in rendering quality and geometric accuracy with an order of magnitude fewer primitives.