FUSER transformer achieves efficient multiview 3D registration with diffusion refinement

Researchers have introduced FUSER, a novel feed-forward transformer designed for multiview 3D registration. This model processes all scans simultaneously in a latent space to directly predict global poses, bypassing the need for computationally intensive pairwise matching. FUSER utilizes a sparse 3D CNN to encode scans and a Geometric Alternating Attention module for efficient reasoning, incorporating 2D attention priors from foundation models. Additionally, FUSER-DF, a diffusion refinement framework, further enhances registration accuracy by denoising estimates in the joint SE(3)$^N$ space. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a novel transformer architecture for 3D registration, potentially improving efficiency and accuracy in applications like robotics and augmented reality.

RANK_REASON This is a research paper introducing a new method for 3D registration.

Read on arXiv cs.CV →

• FUSER
• SE(3)$^N$
• 3DMatch
• ScanNet
• ArkitScenes
• arXiv
• Haobo Jiang

• paper
• other