Images encoded as a few hundred learnable 2D Gaussians, steered by classifier gradients, support a ViT that reaches 76.9% top-1 on ImageNet-1k, close to patch-based ViTs.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
GViT: Representing Images as Gaussians for Visual Recognition
Images encoded as a few hundred learnable 2D Gaussians, steered by classifier gradients, support a ViT that reaches 76.9% top-1 on ImageNet-1k, close to patch-based ViTs.