A two-stage diversity-plus-entropy token selection framework speeds up visual geometry transformers by over 85% on 500-image scenes while preserving baseline accuracy.
VGGT-Long: Chunk it, Loop it, Align it – Pushing VGGT’s limits on kilometer-scale long RGB sequences
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers
A two-stage diversity-plus-entropy token selection framework speeds up visual geometry transformers by over 85% on 500-image scenes while preserving baseline accuracy.