Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers

Researchers propose a two-stage token selection method for visual geometry transformers.

The approach first picks diverse frames, then prunes redundant tokens with attention-entropy-guided, layer-aware sparsification.

It cuts attention cost and reduces the quadratic compute burden in multi-view 3D reconstruction.

On scenes with 500 images, the method reports over 85% speedup while matching or improving quality.