Training a vision transformer with multiple auxiliary tokens and distilling them into a single global token improves self-supervised representation quality with no additional inference cost.
BEit: BERT pre-training of image transformers
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Multi-Token Enhancing for Vision Representation Learning
Training a vision transformer with multiple auxiliary tokens and distilling them into a single global token improves self-supervised representation quality with no additional inference cost.