GViT: Representing Images as Gaussians for Visual Recognition
This work proposes a novel input representation for image classification, which is incremental in improving efficiency but not a broad paradigm shift.
The paper tackles the problem of visual recognition by representing images as compact sets of learnable 2D Gaussians instead of conventional pixel grids, achieving a top-1 accuracy of 76.9% on ImageNet-1k with a ViT-B architecture.
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.