Adding register tokens to Vision Transformers eliminates high-norm background artifacts and raises state-of-the-art performance on dense visual prediction tasks.
ICCV , year=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 1years
2023 1verdicts
UNVERDICTED 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
Vision Transformers Need Registers
Adding register tokens to Vision Transformers eliminates high-norm background artifacts and raises state-of-the-art performance on dense visual prediction tasks.