GViT: Representing Images as Gaussians for Visual Recognition

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Hernandez, Jefferson, He, Ruozhen, Balakrishnan, Guha, Berg, Alexander C., Ordonez, Vicente
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912457315844096
author Hernandez, Jefferson
He, Ruozhen
Balakrishnan, Guha
Berg, Alexander C.
Ordonez, Vicente
author_facet Hernandez, Jefferson
He, Ruozhen
Balakrishnan, Guha
Berg, Alexander C.
Ordonez, Vicente
contents We introduce GVIT, a classification framework that abandons conventional pixel or patch grid input representations in favor of a compact set of learnable 2D Gaussians. Each image is encoded as a few hundred Gaussians whose positions, scales, orientations, colors, and opacities are optimized jointly with a ViT classifier trained on top of these representations. We reuse the classifier gradients as constructive guidance, steering the Gaussians toward class-salient regions while a differentiable renderer optimizes an image reconstruction loss. We demonstrate that by 2D Gaussian input representations coupled with our GVIT guidance, using a relatively standard ViT architecture, closely matches the performance of a traditional patch-based ViT, reaching a 76.9% top-1 accuracy on Imagenet-1k using a ViT-B architecture.
format Preprint
id arxiv_https___arxiv_org_abs_2506_23532
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle GViT: Representing Images as Gaussians for Visual Recognition
Hernandez, Jefferson
He, Ruozhen
Balakrishnan, Guha
Berg, Alexander C.
Ordonez, Vicente
Computer Vision and Pattern Recognition
Machine Learning
We introduce GVIT, a classification framework that abandons conventional pixel or patch grid input representations in favor of a compact set of learnable 2D Gaussians. Each image is encoded as a few hundred Gaussians whose positions, scales, orientations, colors, and opacities are optimized jointly with a ViT classifier trained on top of these representations. We reuse the classifier gradients as constructive guidance, steering the Gaussians toward class-salient regions while a differentiable renderer optimizes an image reconstruction loss. We demonstrate that by 2D Gaussian input representations coupled with our GVIT guidance, using a relatively standard ViT architecture, closely matches the performance of a traditional patch-based ViT, reaching a 76.9% top-1 accuracy on Imagenet-1k using a ViT-B architecture.
title GViT: Representing Images as Gaussians for Visual Recognition
topic Computer Vision and Pattern Recognition
Machine Learning
url https://arxiv.org/abs/2506.23532