A no-fine-tune token filtering module placed before a vision transformer encoder keeps static high-attention tokens and object-region tokens, yielding a 2.8x inference speedup on SigLIP retrieval with negligible recall loss.
An image is worth 16x16 words: Transformers for image recognition at scale
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
method 1
citation-polarity summary
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1roles
method 1polarities
use method 1representative citing papers
citing papers explorer
-
Speed-up of Vision Transformer Models by Attention-aware Token Filtering
A no-fine-tune token filtering module placed before a vision transformer encoder keeps static high-attention tokens and object-region tokens, yielding a 2.8x inference speedup on SigLIP retrieval with negligible recall loss.